Probabilities Are Not Enough: Formal Controller Synthesis for Stochastic Dynamical Models with Epistemic Uncertainty

The study on epistemic uncertainty due to the lack of knowledge has received increasing attention in risk assessment, reliability analysis, decision making, and design optimization. Different theories have been applied to model and quantify epistemic uncertainty. Research on sensitivity analysis for epistemic uncertainty has also been initialized. Sensitivity analysis can identify the contributions of individual input variables with epistemic uncertainty to the model output. It then helps guide the collection of more information to reduce the effect of epistemic uncertainty. In this paper, an effective sensitivity analysis method for epistemic uncertainty is proposed when both epistemic and aleatory uncertainties exist in model inputs. This method employs the unified uncertainty analysis framework to calculate the plausibility and belief measures. The gap between belief and plausibility measures is used as an indicator of the effect of epistemic uncertainty on the model output. The Kolmogorov-Smirnov distance between the two measures is used to quantify the main effect and the total effect of each independent variable with epistemic uncertainty. By the Kolmogorov-Smirnov distance, the importance of each variable is ranked. The feasibility and effectiveness of the proposed method is demonstrated with two engineering examples.

An honest communication of uncertainty about quantities of interest enhances transparency in scientific assessments. To support this communication, risk assessors should choose appropriate ways to evaluate and characterize epistemic uncertainty. A full treatment of uncertainty requires methods that distinguish aleatory from epistemic uncertainty. Quantitative expressions for epistemic uncertainty are advantageous in scientific assessments because they are nonambiguous and enable individual uncertainties to be characterized and combined in a systematic way. Since 2019, the European Food Safety Authority (EFSA) recommends assessors to express epistemic uncertainty in conclusions of scientific assessments quantitatively by subjective probability. A subjective probability can be used to represent an expert judgment, which may or may not be updated using Bayes's rule to integrate evidence available for the assessment and could be either precise or approximate. Approximate (or bounded) probabilities may be enough for decision making and allow experts to reach agreement on certainty when they struggle to specify precise subjective probabilities. The difference between the lower and upper bound on a subjective probability can also be used to reflect someone's strength of knowledge. In this article, we demonstrate how to quantify uncertainty by bounded probability, and explicitly distinguish between epistemic and aleatory uncertainty, by means of robust Bayesian analysis, including standard Bayesian analysis through precise probability as a special case. For illustration, the two analyses are applied to an intake assessment.

The design of engineering systems is often based on analysis models with substantial errors in predicting failures, that is epistemic uncertainty. The epistemic uncertainty is reduced by post design tests, and the safety of unsafe designs restored by redesign. When this process of design, test and redesign is to be based on probabilistic analysis, there is some controversy on whether uncertainty associated with variability (aleatory uncertainty) should be treated differently than the epistemic uncertainty. In this paper we compare several approaches to design and redesign and treatments of the epistemic uncertainties. These include safety factors, probabilistic approach disregarding redesign and regarding redesign, treating epistemic uncertainty and aleatory uncertainty the same, and more conservative treatment of the epistemic uncertainty. We demonstrate that the proposed approach can allow tradeoff of system performance against development cost (probability of redesign), while a standard reliability based design optimization, which does not take into account future redesign, provides only a single point on the tradeoff curve. We also show that the tradeoff can be achieved even with the traditional safety factor approach, without any probabilistic optimization. Furthermore, we investigate different treatments of epistemic error for probability of failure calculation. We find that it is possible to design to the 95th percentile of the probability of failure with modest mass penalty compared to treating epistemic and aleatory uncertainty alike.

Risk plays an important role in the decision making process. To assess the severity and likelihood of impairment to human health from exposure to a substance or activity that under plausible circumstances can cause harm to human health is the main purpose of risk assessment. It is important to know the nature and characteristic of all available information, data or model parameters which are more generally tainted with aleatory and epistemic uncertainty or both type of uncertainty. In some situation, model parameters are affected by aleatory uncertainty and simultaneously other some parameters are affected by epistemic uncertainty, how far computation of the risk is concern, two ways to deal with the situation either transform all the uncertainties to one type of format or need for joint propagation of uncertainties. In this paper, an effort has been made to combine probability distributions, generalized fuzzy numbers, normal fuzzy numbers and generalized interval valued fuzzy numbers within the same framework.

In practical engineering design, most data sets for system uncertainties are insufficiently sampled from un- known statistical distributions, known as epistemic uncer- tainty. Existing methods in uncertainty-based design optimization have difficulty in handling both aleatory and epistemic uncertainties. To tackle design problems engaging both epistemic and aleatory uncertainties, reliability-based design optimization (RBDO) is integrated with Bayes theorem. It is referred to as Bayesian RBDO. However, Bayesian RBDO becomes extremely expensive when employing the first- or second-order reliability method (FORM/SORM) for reliability predictions. Thus, this paper proposes development of Bayesian RBDO methodology and its integration to a numerical solver, the eigenvector dimension reduction (EDR) method, for Bayesian reliabil- ity analysis. The EDR method takes a sensitivity-free approach for reliability analysis so that it is very efficient and accurate compared with other reliability methods such as FORM/SORM. Efficiency and accuracy of the Bayesian RBDO process are substantially improved after this integration.

Ship design optimization with mixed uncertainty based on evidence theory

The Probabilistic Seismic Demand Analysis (PSDA) which is frequently implemented in the first generation performance-based earthquake engineering quantifies seismic behavior of a structure by computing mean annual frequency of exceeding a specific value of a desired demand parameter given all anticipated earthquakes. This framework, based on the total probability integration formula, provides a technical basis on which aleatory uncertainties, uncertainties originated due to inherent randomness of the phenomena, are explicitly addressed. However, variability in the mean value of different model parameters, referred to as epistemic uncertainties and mainly due the finite-sample size of observations, is neglected. In this study, as an alternative to total probability integration, a reliability-based formulation tailored to effortlessly reflect both aleatory and epistemic uncertainties is put-forward to perform unified PSDA. Next, as an application of the proposed methodology, a reliability-based seismic demand curve of a 4-story example building is developed. Results demonstrate that the Second-Order Reliability Method (SORM) and important sampling method (ISM) along with multi-step Monte Carlo simulation (MSMCS) methods are appropriate candidates for computing reliability-based PSDA with differentiable and nondifferentiable performance functions, respectively.

A novel method for importance measure analysis in the presence of epistemic and aleatory uncertainties

Aiming at providing more accurate information for the first restoration, overhaul and maintenance of airborne products, this paper gives a calendar life assessment method for airborne products considering epistemic and aleatory uncertainty on the basis of investigation and analysis for the factors which affect the uncertainty of calendar life assessment. The uncertainty of calendar life is modeled as two types: epistemic uncertainty and aleatory uncertainty. a aleatory uncertainty model of calendar life based on interval analysis and blind number theory is given with parameter sensitivity analysis. Considering the designer's problem of incomplete cognition of failure mechanism, a measurement method of epistemic uncertainty is given based on the concept of confidence reliability.

Incremental dynamic analysis (IDA) has been extended by introducing a set of structural models in addition to the set of ground motion records which is employed in IDA analysis in order to capture record‐to‐record variability. The set of structural models reflects epistemic (modeling) uncertainties, and is determined by utilizing the latin hypercube sampling (LHS) method. The effects of both aleatory and epistemic uncertainty on seismic response parameters are therefore considered in extended IDA analysis. The proposed method has been applied to an example of the four‐storey‐reinforced concrete frame, for which pseudo‐dynamic tests were performed at the ELSA Laboratory, Ispra. The influence of epistemic uncertainty on the seismic response parameters is presented in terms of summarized IDA curves and dispersion measures. The results of extended IDA analysis are compared with the results of IDA analysis, and the sensitivity of the seismic response parameters to the input random variable using the LHS method is discussed. It is shown that epistemic uncertainty does not have significant influence on the seismic response parameters in the range far from collapse, but could have a significant influence on collapse capacity. Copyright © 2008 John Wiley & Sons, Ltd.

Key Points Methods are provided to quantify information and uncertainty in data Best achievable model performance for a given data set is estimated Model adequacy is characterized in terms of Aleatory and Epistemic Uncertainties

An approach to combining unreliable pieces of evidence and their propagation in a system response analysis

Pressure systems contain hazardous fluids within industrial processes. Inspection plays a vital role in managing the reliability of these safety-critical systems. It allows engineers to identify, characterize, and manage degradation of pressure vessels, piping, and associated equipment to prevent failure and the associated consequences on people and the environment. Mixed uncertainty can affect decision-making in at least three distinct aspects of inspection practice: inspection planning, inspection data analysis, and integrity assessment. Despite this, the inspection engineering discipline lacks methodologies for handling both aleatory and epistemic uncertainties within analyses, which could be expensively misleading. This paper demonstrates the benefits of applying mixed uncertainty quantification and analysis techniques to pressure vessel inspection and integrity assessment through a worked example, which shows how the epistemic and aleatory uncertainty in inspection data can be represented using an imprecise probability approach. The limitations of empirical data are shown to pose challenges to implementing these techniques in practice, and so practical requirements for a framework for implementing uncertainty analysis methods in inspection are proposed. These include, for example, the ability to generate meaningful yet conservative results from even a limited amount of poor-quality data, while allowing results to be bounded more narrowly as more data is collected, findings from better data are pooled, or engineering judgment and assumptions are applied.

Uncertainty in structural dynamics: Experimental validation of a Wishart random matrix model

We expand on existing probabilistic tsunami hazard analysis (PTHA) approaches by presenting a methodology wherein the epistemic uncertainties associated with tsunami timing, generation, and propagation are separated from aleatory uncertainties. The contributions of epistemic and aleatory uncertainties to the overall uncertainty in tsunami hazard are evaluated through a two-level Monte Carlo analysis. The total epistemic uncertainty contribution defines the variance of the tsunami inundation hazard distribution. Through Bayesian inference, the available catalog of tsunami runup data is used to refine the hazard estimate, where the prior distribution is weighted by epistemic uncertainties. The general PTHA framework is further modified to include a hazard function accounting for peak water elevation, velocity, duration of inundation, and warning time. We carry out a formal sensitivity analysis of the tsunami generation and propagation on the feasible model parameter space using the multi-objective generalized sensitivity analysis algorithm. Parameters showing a significant influence on tsunami hazard include the tidal elevation, moment magnitude (M w), the Gutenberg–Richter (G–R) distribution β parameter, epicenter location, and the rake angle. Parameters showing minimal influence on tsunami hazard include the G–R maximum seismic moment magnitude (M max) parameter, strike angle, dip angle, seismic depth and Manning’s roughness. As a proof of concept, the methodology is applied to a case study of tsunami hazard posed to National Oceanic and Atmospheric Administration Station 9411406 (Oil Platform Harvest, CA, USA) by seismic sources along the Aleutian Islands, AK.