Willy Aspinall: Estimating incredibly small event probabilities with expert judgement, Bayes Nets and Importance Sampling: present-day likelihood of a repeat super-eruption of Aso volcano, Japan.

Addressing a volcano hazards litigation issue in a country steeped, almost exclusively, in deterministic decision-making, and prompted indirectly by the 2011 Fukushima Daiichi nuclear accident, we used expert judgements from two separate groups of experts to inform a Bayes Net (BN) uncertainty analysis, aimed at quantifying the infinitesimally small probability of an imagined, near-incredible eruption scenario. The presentation will outline a simplified version of the multidimensional nature of volcanic eruption processes, noting that information available to inform the BN model are both sparse and very uncertain. Apparently conflicting judgements arose within both groups, and the way these dichotomies were treated pragmatically in the BN analysis will be described. A particular challenge was that some of the elicited judgements were provided only as single-point variable values, not as the usual three distribution-defining quantiles; this led on to difficulties representing group-wise Equal Weights Decision Maker solutions for these point-value target items. Further reservations arose when adopting the new, and seemingly flexible, bounded/partly bounded/unbounded options offered by the Metalogistic distribution formulation. Notwithstanding these ostensible shortcomings, Importance Sampling with UNINET allowed us to quantify the vanishingly small target probability; in essence, we show that - when society requires them to do so - scientists need not evade computing occurrence probability distributions for rare, extreme hazards. Sensitivity tests on alternative scenarios and inter-comparison of expert groups’ judgements allow partial appraisals of BN model adequacy to be contextualised in relation to the intended use of the analysis.