Abstract: Some probabilistic safety assessment models based on Bayesian networks have been recommended recently for safety analysis of highways and roads. These methods provide a very natural and powerful alternative to traditional approaches, such as fault and event tree based methods. In this article, we present several new and original contributions to complement the inference engine tools of these models to provide new and relevant information about safety and backward analysis on one hand, and to learn the complex multidimensional joint probabilities of all variables, on the other hand. More precisely, we show how standard tools combined with the partitioning technique can be used in new ways to solve three relevant problems (1) to prognosticate the most probable combinations of variables leading to incidents, (2) to perform a backward analysis to identify the causes of accidents, and (3) to learn the model parameters using Bayesian conjugate methods (categorical and Dirichlet families). Finally, some real examples of applications are used to illustrate the proposed methods.

Autoría

Descargar referencia