Comparison of the learning algorithms for evidence-based BBN modeling A case study on ship grounding accidents

The ship grounding accidents account for more than 50 percent of all maritime accidents in the Gulf of Finland. In other sea areas this number may change, however it is still predominant, thus this type of accident poses certain risks to the maritime transportation systems all around the world. These risks can be evaluated utilizing numerous models through the evaluation of probability and consequences of an accident. However when comes to the next step, namely risk management, these models are not suitable as in most cases they can handle only pre-defined solutions, though the systematic and optimal decision making of the risk in the transportation system cannot be made. This paper proposes a methodology for evidence-based BBN risk modeling which makes an attempt to fill this gap. The model gathers data coming from various sources on the grounding accidents that have happened in the Gulf of Finland, organized in the form of chain of events. Thereby, the accident reports of the ship grounding accidents have been used as well as the experts’ knowledge elicitation sessions have been held to structure the model. The paper then compares the available algorithms for learning the structure of a BBN from the datasets that is extracted from the accident reports. Although “Bayesian Search” algorithm is shown to work better for the studied problem and with the utilized data, it is recommended not to rely only on a single algorithm while making BBN risk models based on the evidence extracted from the accident reports. the accident is the key question for the decision makers and the risk managers. Traditionally, risk managers rely on the risk models that model the desired systems and mimic their behaviors under different conditions to define what kind of barriers or RCO should be defined and where they should be placed to protect the system from the accident. Nevertheless, the effectiveness of the implemented RCO highly depends on the accuracy and reliability of the utilized risk models for analyzing the system. Therefore, in order to effectively mitigate the risk, the implemented risk model should firstly reflect the knowledge on the analyzed system with satisfying accuracy (Aven, 2013), and secondly be able to suggest the feasible and meaningful measures for lessening the involved risk (Montewka et al., 2013a). Bayesian Belief Network (BBN) models are considered as proper tools that are able to suggest