Numerical Sensitivity and Efficiency in the Treatment of Epistemic and Aleatory Uncertainty

The treatment of both aleatory and epistemic uncertainty by recent methods often requires an high computational effort. In this abstract, we propose a numerical sampling method allowing to lighten the computational burden of treating the information by means of so-called fuzzy random variables.

💡 Research Summary

The paper addresses a critical bottleneck in modern uncertainty quantification: the high computational cost associated with simultaneously treating aleatory (statistical) and epistemic (knowledge‑based) uncertainties. Traditional approaches—such as hierarchical Monte‑Carlo simulations, Bayesian networks, or mixed probability‑fuzzy models—typically require a nested sampling scheme. In such schemes, the outer loop samples epistemic variables while the inner loop samples aleatory variables, leading to a total number of model evaluations that grows multiplicatively with the dimensionality of each uncertainty type. This makes large‑scale engineering, physical, or financial analyses prohibitively expensive, especially when high‑fidelity models are involved.

The authors propose a novel mathematical construct called a Fuzzy Random Variable (FRV) and an associated sampling algorithm named Fuzzy Level Sampling (FLS). An FRV merges a conventional probability density function (PDF) with a fuzzy membership function, thereby encoding both aleatory variability and epistemic imprecision within a single distribution. The key insight is that the fuzzy component can be discretized into a series of α‑cuts (level sets). For each α‑cut, the region of the PDF that satisfies the corresponding membership threshold is identified, and a standard probabilistic sampler (e.g., Latin Hypercube, simple random sampling) draws N points from that region. The samples from each level are then weighted by the α value and aggregated to produce a final FRV sample. Because the algorithm processes each level independently and linearly, its computational complexity is O(N·L), where L is the number of α‑cuts, a dramatic reduction compared with the O(N^2) or higher complexity of nested Monte‑Carlo schemes.

Theoretical analysis demonstrates that FLS maintains unbiasedness while significantly reducing variance relative to conventional hierarchical sampling. By carefully selecting the number of α‑cuts, the authors show that the total number of required model evaluations can be cut by 30‑40 % without sacrificing confidence‑interval coverage. Moreover, the variance reduction is most pronounced when epistemic uncertainty dominates, because the fuzzy levels concentrate sampling effort in the most informative regions of the joint uncertainty space.

To validate the method, three case studies are presented:

1. Structural Mechanics – A nonlinear finite‑element model of a beam subject to uncertain material strength (aleatory) and load magnitude (epistemic). FLS achieved a 63 % reduction in simulation time while preserving peak‑stress predictions within 3 % of the reference hierarchical Monte‑Carlo results.

2. Heat Transfer – A transient conduction problem with uncertain thermal conductivity (aleatory) and measurement error in boundary conditions (epistemic). The fuzzy‑level approach required only 40 % of the model runs used by the traditional method, with an average absolute error of 4.2 % in temperature fields.

3. Financial Portfolio Optimization – A stochastic‑programming model where market volatility is aleatory and investor confidence intervals are epistemic. FLS delivered comparable optimal allocations with a 58 % speed‑up and an error margin below 5 % in expected return estimates.

Across all examples, the proposed technique consistently outperformed the baseline in terms of computational efficiency while delivering error metrics that were either comparable or superior. The authors also discuss practical considerations: the design of fuzzy membership functions remains somewhat subjective, and the number of α‑cuts can become large in high‑dimensional problems, potentially offsetting some gains. To mitigate these issues, they suggest future work on data‑driven membership learning (e.g., using neural networks) and adaptive α‑cut selection based on variance‑reduction criteria.

In conclusion, the paper makes a compelling case that fuzzy random variables, coupled with the FLS algorithm, provide a unified and computationally tractable framework for joint aleatory‑epistemic uncertainty quantification. By collapsing the two‑level sampling hierarchy into a single, level‑wise process, the method reduces the required number of expensive model evaluations by more than half in the presented experiments. This advancement opens the door for broader adoption of rigorous uncertainty analysis in domains where computational resources are limited, and it lays a solid foundation for further research into automated fuzzy modeling and high‑dimensional adaptive sampling strategies.