The session of Bayesian Reasoning contains three presentations. The first presenter in this session was Johan Kwisthout from Radboud University Nijmegen with the paper entitled “Most Frugal Explanations: Occam’s Razor Applied to Bayesian Abduction”. Johan Kwisthout discussed whether the traditional best explanation which is the most probable explanation is indeed the “best” one. He proposes a different “best explanation” namely the most frugal explanation. They start from the idea that when less variables are needed for a decision this explanation is considered as the best explanation (this is Occam’s Razor). He shows that this new notion of best explanation is even a harder problem then the traditional one, but that the Most Frugal Explanation can be tractably approximated under situational constraints that are more realistic then the ones that are needed for the traditional (MAP) tractable.
Sjoerd Timmer presented their paper entitled “Inference and Attack in Bayesian Networks”. In the legal domain is nowadays a need for methods to assess the probabilistic evidence (for instance DNA matching results). This type of evidence needs to be included in the whole reasoning process of the case beside the usual used argumentation models which are more natural to follow for lawyers and judges. They propose a method to extract inference rules and undercutters from Bayesian networks such these can be used in the construction of arguments. Very interesting to see that they make a connection to those rather different types of models: Bayesian networks and argumentation theory.
The last presentation was given by Janneke Bolt. She presented the paper entitled ”Towards Uncertainty Analysis of Bayesian Networks” from her colleagues in Utrecht very well.
The easy part of the modeling of Bayesian Network is the modeling of the graph. The difficult part is to get the right parameters values. For analyzing the inaccuracies in the parameter probabilities usual a sensitive analysis is performed. However a sensitivity analysis is rather difficult by more parameters, therefore they propose another method to test the robustness of the network parameters probabilities namely by an uncertainty analysis. The research was in a preliminary stage, but the first results on their small fictitious example seems promising.
