Short communication A virtual experimental technique for data collection for a Bayesian network approach to human reliability analysis Mashrura Musharraf, David Bradbury-Squires, Faisal Khan n , Brian Veitch, Scott MacKinnon, Syed Imtiaz Faculty of Engineering & Applied Science, School of Human Kinetics and Recreation, Memorial University of Newfoundland, Newfoundland and Labrador, St John's, Canada A1B 3X5 article info Article history: Received 18 October 2013 Received in revised form 10 June 2014 Accepted 27 June 2014 Available online 5 July 2014 Keywords: Human factor Human reliability analysis Bayesian network abstract Bayesian network (BN) is a powerful tool for human reliability analysis (HRA) as it can characterize the dependency among different human performance shaping factors (PSFs) and associated actions. It can also quantify the importance of different PSFs that may cause a human error. Data required to fully quantify BN for HRA in offshore emergency situations are not readily available. For many situations, there is little or no appropriate data. This presents significant challenges to assign the prior and conditional probabilities that are required by the BN approach. To handle the data scarcity problem, this paper presents a data collection methodology using a virtual environment for a simplified BN model of offshore emergency evacuation. A two-level, three-factor experiment is used to collect human performance data under different mustering conditions. Collected data are integrated in the BN model and results are compared with a previous study. The work demonstrates that the BN model can assess the human failure likelihood effectively. Besides, the BN model provides the opportunities to incorporate new evidence and handle complex interactions among PSFs and associated actions. & 2014 Elsevier Ltd. All rights reserved. 1. Introduction Bayesian networks (BNs) are acyclic directed graphs modeling probabilistic dependencies and interdependencies among variables [11]. The graphical part of BN reflects the causal relationship of the variables under consideration. The interactions among these vari- ables are quantified by conditional probabilities. BNs have proved to be a powerful tool for human reliability analysis (HRA) [12]. The BN has the capability to consider dependency among human perfor- mance shaping factors (PSFs) [2] and associated actions [13]. Also, human reliability calculated using a BN model can be updated each time new evidence or information is available [9]. A major difficulty in applying BN to practical problems is to obtain the numerical parameters that are needed to fully quantify a BN [10]. The conditional probability distribution table (CPT) for a binary variable with n binary predecessors in a BN requires specification of 2 n independent parameters. For a sufficiently large n, eliciting 2 n parameters is difficult. This problem is severe in the case of HRA for emergency scenarios as human performance data are not readily available. Though expert judgment techniques have been used as a solution to the data scarcity problem, collecting judgment from domain experts for 2 n parameters can be prohibitively cumbersome when n is large. Groth and Mosleh [6] describes a methodology to combine multiple sources of empirical data (the Human Events Repository Analysis (HERA) database and worksheets from an applica- tion of the Information-Decision-Action (IDA) model) to develop a data-informed Bayesian (belief) network. The current paper presents a way to collect human performance data using a virtual experimental technique to deal with the data scarcity problem. The collected data are integrated in the BN model to determine the human error probability. The outcomes are then compared with a previous study [3] that uses the same data but a different methodologies. A BN model is developed to observe the effect on human performance in offshore evacuation of three PSFs: training, visibility and complexity. Three different responses were measured: time to evacuation, backtracking time, and exposure to hazards. Data needed to fully quantify this BN model (2 n combinations of n factors) were collected by a two level (assuming all factors are binary) n factor experiment in a virtual environment (VE). An overview of the VE used in the experiment is given in Section 2. The detail of the experimental design is given in Section 3. Section 4 describes the data collection process. Section 5 illustrates integration of collected data in the BNs. Results are discussed and compared with the previous study in Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/ress Reliability Engineering and System Safety http://dx.doi.org/10.1016/j.ress.2014.06.016 0951-8320/& 2014 Elsevier Ltd. All rights reserved. n Corresponding author. Reliability Engineering and System Safety 132 (2014) 1–8