In the Uncertainty in AI session three papers were presented. Two papers were wholly or partly motivated by legal applications. Remi Wieten from Utrecht university addressed the problem that Bayesian networks, while increasingly popular in forensic science, are poorly understood by legal experts. This problem makes it hard to build reliable Bayesian networks for crime investigation or for evidential reasoning in criminal cases. Wieten further developed the idea that legal experts can express their evidential knowledge in terms of arguments (a style of reasoning with which they are familiar), which are then automatically converted into constraints on the design of a Bayesian network.
Joost Vennekens of the KU Leuven presented a joint paper with Sander Beckers on formalizing the concept of actual causation, where the problem is to define when event X is deemed to have caused event Y in the context of a particular story. This is a problem that often arises in legal cases. After criticizing David Lewis’ definition of actual causation as the transitivity of counterfactual dependency, Vennekens presented an alternative approach based on the ideas that counterfactual dependency is a sufficient but not necessary condition for actual causation, and that actual causation is transitive only insofar as it does not violate asymmetry (X cannot have caused Y if we would also have considered not-X as a cause of Y).
Johan Quisthout from the Donders Institute in Nijmegen addressed complexity issues of approximate inference in Bayesian networks. One motivation of this work is the Bayesian Brain hypothesis from cognitive science, according to which the human brain carries out or at least approximates Bayesian updating. This raises the issue of how humans can do this efficiently given that Bayesian inference is intractable in general. Using the formal framework of so-called fixed-error randomized tractability, Kwisthout presented a number of positive and negative tractability results.
