Machine learning for human reliability and interaction with intelligent systems

The present research proposes to develop an innovative approach to construct an calibrate a model for calculating the human error probability using data extracted from existing major accident investigation reports. This approach has the potential to provide data that depict contexts and scenarios not fully achieved by simulator, near-misses and expert elicitation data.
Credal Networks and imprecise probability and text analytics tools will be adopted to model the relationship between performance shaping factors and human errors.
Credal Networks are extension of Bayesian Networks. They offer the possibility of identifying the reasons for the result of interest rather than providing single numerical values and they can deal with different representation of uncertainty by providing confidence bounds to the results. This is particular important in the definition of the Conditional probability tables required by the model and the possibility of discriminating a lack of data from no events. For instance, observing 1 event over 2 or 50 over 100 produces the same probability but the confidence in the number need to be included in the analysis. Therefore, confidence boxes will be used to characterise such imprecision. Graphical models are also more accessible to different research disciplines, facilitating multi-disciplinary risk analysis.
Finally, the proposed approach will be applied to analyse and model human reliability during emergency and recovery situation, as they are considered to require significant human intervention to re-establish normal operation. The objective functions will be not only the safety of the workers and adjacent communities but also the business continuity, as interconnected critical infrastructure can affect essential services provided to population resulting in socioeconomic losses over time.