Nicolas Kühn, Institut für Geowissenschaften, Universität Potsdam Title: Modeling the Joint Probability of Earthquake, Site and Ground-Motion Parameters Using Bayesian Networks Time: February 11?, 15:00 Place: Building 28, Room 2.123, Campus Golm, Universitaet Potsdam Abstract: Empirical ground-motion models for use in seismic hazard analysis are commonly described by regression models, where the ground-motion parameter is assumed to be dependent on some earthquake- and site-specific parameters such as magnitude, distance or local vs30. In regression analysis only the target is treated as a random variable, while the predictors are not; they are implicitly assumed to be complete and error-free, which is not the case for magnitudes or distances in earthquake catalogs. However, in research areas such as machine learning or artificial intelligence techniques to overcome these issues exist. Borrowing from these fields, we present a novel multivariate approach to ground-motion estimation by means of the Bayesian network (BN) formalism. This elegant and intuitively appealing framework allows for reasoning under uncertainty by modeling directly the joint probability distribution of all variables, while at the same time offering explicit insight into the probabilistic relationships between variables. The formalism provides us with efficient methods for computing any marginal or conditional distribution of any subset of variables. In particular, if some earthquake- or site-related parameters are unknown, the distribution of the ground motion parameter of interest can still be calculated. In this case, the associated uncertainty is incorporated in the model framework. Here, we explore the use of BNs in the development of ground-motion models. Therefore, we construct BNs for both a synthetic and the NGA dataset, the most comprehensive strong ground motion dataset currently available. The analysis shows that BNs are able to capture the probabilistic dependencies between the different variables of interest. Comparison of the learned BN with the NGA model of Boore and Atkinson (2008) shows a reasonable agreement in distance and magnitude ranges with good data coverage.