Advances in dataset analysis techniques have enabled more sophisticated approaches to analyzing and characterizing training data instances, often categorizing data based on attributes such as "difficulty".In this work, we introduce RepMatch, a novel method that characterizes data through the lens of similarity.RepMatch quantifies the similarity between subsets of training instances by comparing the knowledge encoded in models trained on them, overcoming the limitations of existing analysis methods that focus solely on individual instances and are restricted to within-dataset analysis.Our framework allows for a broader evaluation, enabling similarity comparisons across arbitrary subsets of instances, supporting both dataset-to-dataset and instance-to-dataset analyses.We validate the effectiveness of Rep-Match across multiple NLP tasks, datasets, and models.Through extensive experimentation, we demonstrate that RepMatch can effectively compare datasets, identify more representative subsets of a dataset (that lead to better performance than randomly selected subsets of equivalent size), and uncover heuristics underlying the construction of some challenge datasets.

The concept of physics of failure (PoF) emerged from fields, such as damage mechanics, corrosion engineering, and fracture mechanics. Recognizing the increasing complexity of military equipment and the rise in observed failures, the Rome Air Development Center (RADC), which preceded the U.S. Air Force Rome Laboratory, introduced the PoF concept in 1961. One of the very early PoF models was presented by researchers from Bell Labs, who justified using the kinetic theory's interpretation of the Arrhenius equation to assess the temperature-induced aging of semiconductor devices [1]. During the early 1960s, the RADC and Armor Research Foundation of the Illinois Institute of Technology organized symposia to introduce and explain PoF concepts and methods. These symposia paved the way for numerous research activities and events that continue to this day.

This paper presents a new method for projecting the reliability growth of a complex continuously operating system. The model allows for arbitrary corrective action strategies, and it differs from other models of this type by using all available data rather than failure mode first occurrence times only. It also differs from other reliability growth projection models in that it provides a complete inference framework via the posterior distribution on the system failure intensity. A unique feature of this approach relative to other Bayesian techniques is the analytic expression for the failure intensity contribution from unobserved failure modes. Expressions for estimating the initial failure intensity, growth potential failure intensity, and the cumulative number of failure modes expected in future testing are also developed. Extensions to the basic framework are also developed. The first accounts for multiple systems under test, and the second develops the posterior distribution while allowing for uncertainty on the Fix Effectiveness Factor values that are assessed. Two separate goodness-of-fit procedures are presented for assessing the appropriateness of the underlying model assumptions.

We propose a robust optimization framework to deal with uncertain component reliabilities in redundancy allocation problems in series-parallel systems. The proposed models are based on linearized versions of standard mixed integer nonlinear programming (MINLP) formulations of these problems. We extend the linearized models to address uncertainty by assuming that the component reliabilities belong to a budgeted uncertainty set, and develop robust counterpart models. A key challenge is that, because the models involve nonlinear functions of the uncertain data, classical robust optimization approaches cannot apply directly to construct their robust optimization counterparts. We exploit problem structure to develop robust counterparts and exact solution methods, and present computational results demonstrating their performance.

This paper studies a double excited 3-Degree Of Freedom (3-DOF) actuator using a nonlinear equivalent magnetic circuit to perform its characteristic analysis in a short time. To investigate the exact characteristic analysis, the nonlinear feature was considered for the magnetic saturation, leakage flux, airgap flux, and fringing effect of a double excited 3-DOF actuator. The nonlinear equivalent magnetic circuit result is compared with the 3D Finite Element Method (FEM) and experiment results to demonstrate its validity.

The objective of this research is to develop a practical research and development (R&D) portfolio selection model that addresses the effective R&D project valuation issue, while tackling R&D uncertainty in portfolio optimization. Fuzzy set theory is employed to capture and model the uncertain project information. To evade the well-known complexities of fuzzy real option valuation, the recently developed fuzzy pay-off method is used to more effectively valuate R&D projects. The resulting problem is formulated as a fuzzy zero-one integer programming model that handles uncertainty of input data in order to determine the optimal portfolio. Two satisfaction measures, which are based on possibility and necessity theory, are employed to solve the resulting fuzzy model. An example is presented to compare these measures, as well as to illustrate the simplicity and applicability of the proposed approach.

We propose a robust deviation framework to deal with uncertain component reliabilities in the constrained redundancy optimization problem (CROP) in series-parallel reliability systems. The proposed model is based on a linearized binary version of standard nonlinear integer programming formulations of this problem. We extend the linearized model to address uncertainty by assuming that the component reliabilities belong to an interval uncertainty set, where only upper and lower bounds are known for each component reliability, and develop a Min-Max regret model to handle data uncertainty. A key challenge is that, because the deterministic model involves nonlinear functions of the uncertain data, classical robust deviation approaches cannot be applied directly to find robust solutions. We exploit problem structures to develop four exact solution methods, and present computational results demonstrating their performance.

In this paper, fuzzy linear regression models with fuzzy/crisp output, fuzzy/crisp input are considered. In this regard, we define risk-neutral, risk-averse and risk-seeking fuzzy linear regression models. In order to do that, two equality indices are applied to express the degree of equality between a pair of fuzzy numbers. We also develop three mathematical models to obtain the parameters of fuzzy linear regression models. Minimizing the difference between the total spread of the observed and estimated values is the objective of these models. The advantage of our proposed models is the simplicity in programming and computation.