Handling the uncertainty with confidence in human reliability analysis

Most of the attempts aimed at substituting expert-driven human reliability assessment methods with empirical data-driven techniques have failed due to the high uncertainty of human reliability databases and limitations of traditional probabilistic tools to deal with it. Although recent research suggests Bayesian and credal networks could be a more suitable approach to model human reliability data, such analyses implies the need for the assessment of a conditional probability distribution for each variable – requiring a much larger amount of data than other traditional tools. Therefore, 'the problem of sparse data' continues to play a crucial role in hindering the feasibility and credibility of human reliability analysis. This has fuelled research aiming at tackling data scarcity through the use of expert elicitation and, more recently, of imprecise probability. In addition to issues inherent to the nature of the available data, some modelling procedures such as normalisation have the potential to implicitly affect the degree of knowledge carried by such data, resulting in loss of reliability. For instance, our confidence about the probability of an event that has been observed in only one of ten trials (1/10) is not the same as that of an event observed to occur ten times in one hundred trials (10/100). Hence, the output of such a procedure does not carry any information regarding the unevenness of sample sizes. In this paper, we propose to tackle these limitations by using confidence boxes (c-boxes) with credal networks, aiming at providing risk assessors with a rigorous framework for data uncertainty guiding towards more efficient and robust modelling solutions. The approach is tested with a simple model of the causes of fatigue in the work environment.