System reliability analysis for rare events based on improved cross-entropy importance sampling and parallel learning strategy

Stochastic collocation enhanced radial-circumferential importance sampling method for efficiently estimating rare failure probability with high-dimensional inputs and its application in slope structures

Learn safe, robust dynamic treatment regimes (DTRs) from observational trajectories that exhibit treatment selection bias, using an offline reinforcement learning (RL) approach. We propose CQL-RB, which augments Conservative Q-Learning (CQL) with a representation-balancing penalty based on an integral probability metric (IPM) (instantiated as either a maximum mean discrepancy (MMD) or an energy-distance penalty). The penalty aligns latent patient representations across treatment groups to reduce action-conditioned distribution shift while preserving CQL's conservative policy estimation. We evaluate CQL-RB on two clinically realistic simulators: EpiCare (eight environments) and AhnChemo from DTR-Bench, both modeling longitudinal healthcare decisions with binary actions at each stage. To emulate selection bias, we implement clinician-like behavior policies that assign treatment as a function of patient covariates. Baselines include BOWL, ACWL, T-RL, RL-NN, and standard CQL. Outcomes are expected return and adverse-event counts from simulator rollouts; model selection uses weighted importance sampling off-policy evaluation on held-out data. Ablations vary both the IPM weight β and the choice of IPM metric. Across all eight EpiCare environments and the challenging AhnChemo task, CQL-RB with either MMD or energy-distance penalties consistently achieves higher returns than competing methods while yielding lower (or comparable) adverse-event rates. Removing the balancing term degrades both return and safety, confirming its contribution. Performance is robust for moderate penalty weights (e.g., β∈{1,10,100}), with degradation only at overly large values (e.g., β≥1000 for MMD or β=10000 for energy distance). Representation balancing materially strengthens conservative offline RL for DTR learning under treatment selection bias. By aligning patient representations without altering CQL's safety mechanics, CQL-RB delivers policies that are both effective (higher returns) and safer (fewer adverse events) in realistic healthcare simulations. These findings underscore the importance of addressing treatment selection bias when learning robust and safe dynamic treatment policies.

Operator-aware shadow importance sampling for accurate fidelity estimation

We develop and evaluate a deep contextual bandit framework for multivariate off-policy evaluation within a controlled simulation-based validation setting. Using real covariate distributions from the Adult, Boston Housing, and Wine Quality datasets, we construct synthetic treatment assignments and multivariate potential outcomes to enable rigorous benchmarking under known data-generating processes. We compare CNN-LSTM, LSTM, and Feed-forward Neural Network (FNN) architectures as nonlinear action-value estimators. To examine representation learning under structured dependence, an AR(1) feature augmentation scheme is employed, while multivariate outcomes are standardized using empirical copula transformations to preserve cross-dimensional dependence. Policy values are estimated using Stabilized Importance Sampling (SIPS) and doubly robust (DR) estimators with bootstrap inference. Although the decision problem is strictly one-step, empirical results indicate that CNN-LSTM architectures provide competitive action-value calibration under temporal augmentation. Across all datasets, the DR estimator demonstrates substantially lower variance and greater stability than SIPS, consistent with its theoretical variance-reduction properties. Diagnostic analyses—including propensity overlap assessment, cumulative oracle regret (with oracle values known by construction), calibration evaluation, and sensitivity analysis—support the reliability of the proposed evaluation framework. Overall, the results demonstrate that combining copula-normalized multivariate outcomes with doubly robust off-policy evaluation yields a statistically principled and variance-efficient approach for offline policy learning in high-dimensional simulated environments.

Training robust neural network by importance sampling and minibatching

The Bayesian analysis of infectious disease surveillance data from multiple locations typically involves building and fitting a spatio-temporal model of how the disease spreads in the structured population. Here we present new generally applicable methodology to perform this task. We introduce a parsimonious representation of seasonality and a biologically informed specification of the outbreak component to avoid parameter identifiability issues. We develop a computationally efficient Bayesian inference methodology for the proposed models, including techniques to detect outbreaks by computing marginal posterior probabilities at each spatial location and time point. We show that it is possible to efficiently integrate out the discrete parameters associated with outbreak states, enabling the use of dynamic Hamiltonian Monte Carlo (HMC) as a complementary alternative to a hybrid Markov chain Monte Carlo (MCMC) algorithm. Furthermore, we introduce a robust Bayesian model comparison framework based on importance sampling to approximate model evidence in high-dimensional space. The performance of our methodology is validated through systematic simulation studies, where simulated outbreaks were successfully detected, and our model comparison strategy demonstrates strong reliability. We also apply our new methodology to monthly incidence data on invasive meningococcal disease from 28 European countries. The results highlight outbreaks across multiple countries and months, with model comparison analysis showing that the new specification outperforms previous approaches. The accompanying software is freely available as a R package at https://github.com/Matthewadeoye/DetectOutbreaks.

Composite power system reliability evaluation using Monte Carlo simulation often suffers from high computational cost due to the difficulty in capturing rare loss-of-load states. To address this challenge, a novel technique called Random Line Importance Sampling integrated with the Cross-Entropy method (RLIS-C) is proposed. Unlike crude Monte Carlo simulation and the conventional cross-entropy method, which operate in a mixed discrete–continuous space, the proposed approach transforms this irregular space into a standard normal space composed of independent standard normal variables. In this transformed space, a random line sampling strategy based on spatial geometry is developed. Furthermore, RLIS constructs a one-dimensional Kriging surrogate-based importance sampling density function for each randomly generated failure line, enabling efficient extraction of rare loss-of-load states on the failure line. In addition, RLIS-C using the cross-entropy method to zoom in on the rare loss-of-load domain is further presented to enhance the capturing efficiency of failure lines. The proposed RLIS-C demonstrates clear efficiency gains over standalone RLIS or cross-entropy approaches. Its effectiveness is validated through reliability evaluations on an illustrative 3-bus system, as well as the IEEE-RTS79, IEEE-RTS96 systems, including their modified versions.

Estimating predictive failure probability and its update under newly available observations by a layered cluster importance sampling algorithm

Spatially distributed infrastructure systems are crucial assets of a community, and thus assessing their risk under seismic hazards is essential. These assessments should be system-wide, regional, and probabilistic to provide comprehensive insights for decision-makers. Monte Carlo simulation is often used owing to its straightforward application but can be computationally inefficient for rare, high-impact events. While importance sampling can improve computational efficiency by focusing on significant outcomes, it may struggle with computing multiple probabilities simultaneously as it is usually optimized for single-event scenarios. To address these issues, we propose a novel approach termed cross-entropy-based “concurrent” adaptive importance sampling (CE-CAIS). This method efficiently samples multistate systems by concurrently estimating multiple system failure probabilities using a single near-optimal importance sampling density. The effectiveness of CE-CAIS is demonstrated using the Sioux Falls traffic network example consisting of 11 system states and a large-scale regional analysis. In the latter, dimensionality reduction techniques, such as principal component analysis, and the central limit theorem are additionally employed to assess the loss of Shelby County, Tennessee, which includes 305,694 buildings. These results highlight the potential of CE-CAIS in risk assessment of complex systems, paving the way for more effective and scalable approaches to evaluating regional impacts of seismic hazards on spatially distributed infrastructure systems.

We study inference and prediction for two populations whose lifetimes follow two-parameter Birnbaum–Saunders distributions under a joint progressive Type-II censoring scheme. We derive the observed-data likelihood and obtain maximum likelihood estimates via an EM algorithm that treats progressively removed lifetimes as missing data. Bayesian inference is developed using importance sampling and a hybrid Gibbs–Metropolis–Hastings sampler, leading to Bayes estimators, credible intervals, and posterior predictive summaries. We further construct prediction intervals for the unobserved lifetimes removed at multiple censoring stages. Monte Carlo experiments under several censoring patterns and parameter configurations compare the frequentist and Bayesian procedures. A tuberculosis survival dataset illustrates model adequacy, parameter estimation, and prediction of removed units under joint progressive censoring.

The coming decade will be crucial for determining the final design and configuration of a global network of next-generation (XG) gravitational-wave detectors, including the Telescope (ET) and Cosmic Explorer (CE). In this study, and for the first time, we assessed the performance of various network configurations using neural posterior estimation (NPE) implemented in Einstein Dingo-IS ---a method based on normalizing flows and importance sampling that enables fast and accurate inference. We focused on a specific science case involving short-duration, massive and high-redshift binary black hole mergers with detector-frame chirp masses $( _ )$ $> 100 _⊙$. These systems encompass early-Universe stellar and primordial black holes, as well as intermediate-mass black hole binaries, for which XG observatories are expected to deliver major discoveries. Validation against standard Bayesian inference demonstrates that NPE robustly reproduces complex and disconnected posterior structures across all network configurations. For a network of two misaligned L-shaped ET detectors (2L MisA), the posterior distributions on luminosity distance can become multimodal and degenerate with the sky position, leading to less precise distance estimates compared to the triangular ET configuration. However, the number of sky-location multimodalities is substantially lower than the eight expected with the triangular ET, resulting in improved sky and volume localization. Adding CE to the network further reduces sky-position degeneracies, and the better performance of the 2L MisA configuration over the triangle remains evident. M d M

This paper applies the reliability-based robust optimization (RBRO) technique to investigate the probabilistic structural design characteristics of the Fairlead Chain Stopper (FCS), a newly developed detachable mooring apparatus for installation on a 10 MW floating offshore wind turbine. The thickness dimensions of the FCS’s major structural members were considered as random design variables, including uncertainties such as manufacturing tolerances. The structural strength performance was defined as a probabilistic constraint function based on the allowable stresses specified by DNV classification rule. The structural strength performance of the FCS was evaluated through finite element analysis (FEA) using design load conditions for moored (LC1, LC2) and towed (LC3) conditions based on DNV classification rules. The RBRO design problem was formulated with weight minimization as the objective function, with probabilistic constraints on strength performance and 3-sigma robustness applied as side constraints. To evaluate reliability analysis methods suitable for probabilistic optimal design, the Mean Value Reliability Method (MVRM) and the Adaptive Importance Sampling Method (AISM) were applied during the RBRO process, and the results were compared and analyzed. The probabilistic optimal design using RBRO exhibited conservative design characteristics compared to the deterministic optimal design, ensuring robustness and reliability. After comprehensively considering the weight reduction rate and numerical computational cost (number of function evaluations), the RBRO method using MVRM was confirmed to be the most reasonable method for the probabilistic optimal structural design of the FCS.

Retrieval-augmented generation (RAG) integrates large language models (LLMs) with retrievers to access external knowledge, improving the factuality of LLM generation in knowledge-grounded tasks. To optimize the RAG performance, most previous work independently fine-tunes the retriever to adapt to frozen LLMs or trains the LLMs to use documents retrieved by off-theshelf retrievers, lacking end-to-end training supervision. Recent work addresses this limitation by jointly training these two components but relies on overly simplifying assumptions of document independence, which has been criticized for being far from real-world scenarios. Thus, effectively optimizing the overall RAG performance remains a critical challenge. We propose a direct retrieval-augmented optimization framework, named DRO, that enables end-to-end training of two key components: (i) a generative knowledge selection model and (ii) an LLM generator. DRO alternates between two phases: (i) document permutation estimation and (ii) re-weighted maximization, progressively improving RAG components through a variational approach. In the estimation step, we treat document permutation as a latent variable and directly estimate its distribution from the selection model by applying an importance sampling strategy. In the maximization step, we calibrate the optimization expectation using importance weights and jointly train the selection model and LLM generator. Our theoretical analysis reveals that DRO is analogous to policy-gradient methods in reinforcement learning. Extensive experiments conducted on five datasets illustrate that DRO outperforms the best baseline with 5%–15% improvements in EM and F1. We also qualitatively analyze the stability, convergence, and variance of DRO.

Collaborative control of multiple surface vessels remains a significant challenge in autonomous maritime operations, particularly within environments characterized by sparse rewards. Conventional Multi-Agent Proximal Policy Optimization (MAPPO) often suffers from inefficient credit assignment and slow convergence in such scenarios. To address these limitations, this paper proposes an enhanced MAPPO framework that integrates a counterfactual baseline—derived from Counterfactual Multi-Agent Policy Gradients (CMAPG)—into the Generalized Advantage Estimation (GAE) formulation. Furthermore, a Prioritized Experience Replay (PER) mechanism with importance sampling is incorporated to improve sample efficiency. The counterfactual baseline is necessary to provide precise, agent-specific learning signals within the on-policy paradigm, directly tackling the credit assignment problem. The PER mechanism, carefully adapted with importance sampling, is essential to break the sample-inefficiency barrier by strategically reusing valuable past experiences without compromising stability. This synergistic approach refines credit assignment by isolating individual contributions and maximizes the utility of valuable historical experiences. Simulation results and comparisons validate the enhanced control performance of the proposed controller.

With the integration of high-penetration power electronics, the dynamic characteristics of modern power systems are jointly dominated by synchronous generators (SGs) and virtual synchronous machines (VSMs). However, the accuracy of these system parameters cannot always be guaranteed in real-world scenarios. To tackle this issue, we propose a robust parameter identification and correction framework based on trajectory sensitivity analysis and Pareto smoothed importance sampling (PSIS). First, to identify the sources of dynamic anomalies, we employ trajectory sensitivity analysis to quantify the impact of parameter variations and screen out key parameters for targeted identification. Subsequently, to utilize the readily available ambient measurements, we incorporate successive variational mode decomposition (SVMD). This method adaptively extracts the dominant variation modes, thereby recovering high-quality data for the identification process. Finally, to circumvent the weight degradation problem inherent in traditional particle filters, we propose a cost-effective PSIS algorithm to obtain the robust posterior distribution of modern system parameters. Simulation results demonstrate the excellent performance of the proposed method. It can not only precisely estimate the key parameters of both SGs and VSMs but also realize the automatic correction of dynamic models under complex operating scenarios.

The goal of prompt nuclear forensics is to determine the characteristics of a nuclear detonation based on the signatures available almost immediately after the explosion. An important characteristic is the reaction time history (RTH), a measure of the device’s rate of neutron multiplication. The RTH can be estimated by observation of the gamma radiation emitted from the detonation, which can be detected directly or observed indirectly as Teller light. Gamma transport simulations used to predict these radiation fields are often modeled stochastically using the Monte Carlo N-Particle (MCNP) code, which can be a computationally demanding task due to the number of particle histories needed to achieve statistical convergence. In an attempt to improve the efficiency of these calculations, we evaluate two variance reduction techniques: Consistent Adjoint-Driven Importance Sampling (CADIS) and Forward-Weighted Consistent Adjoint-Driven Importance Sampling (FW-CADIS). These methods use a deterministically calculated adjoint flux to create weight windows and source biasing that guide MCNP sampling. We study the utility of CADIS and FW-CADIS for their use in MCNP gamma transport for nuclear forensics prediction simulations. The results demonstrate that both CADIS and FW-CADIS improve the accuracy for forensics-focused simulations, with CADIS being most beneficial in direct detection and FW-CADIS being ideal for computing a global Teller light source.

Abstract Neural-network quantum states (NQS) offer a powerful and expressive ansatz for representing quantum many-body wave functions. However, their training via Variational Monte Carlo (VMC) methods remains challenging. It is well known that some scenarios -such as sharply peaked wave functions emerging in quantum chemistry -lead to high-variance gradient estimators hindering the effectiveness of variational optimizations. In this work we investigate a systematic strategy to tackle those sampling issues by means of adaptively tuned importance sampling. Our approach is explicitly designed to target the gradient estimator instead of the loss function and be computationally inexpensive. We benchmarked our approach across the ground-state search of a wide variety of hamiltonians, including frustrated spin systems and ab-initio quantum chemistry. Overall, our approach can reduce the computational cost of vanilla VMC considerably, up to a factor of 100x when targeting highly peaked quantum chemistry wavefunctions.

Sequential multiple importance sampling for multi-modal Bayesian inference

Adaptive sampling algorithms are modern and efficient methods that dynamically adjust the sample size throughout the optimization process. However, they may encounter difficulties in risk-averse settings, particularly due to the challenge of accurately sampling from the tails of the underlying distribution of random inputs. This often leads to a much faster growth of the sample size compared to risk-neutral problems. In this work, we propose a novel adaptive sampling algorithm that adapts both the sample size and the sampling distribution at each iteration. The biasing distributions are constructed on the fly, leveraging a reduced-order model of the objective function to be minimized, and are designed to oversample a so-called risk region. As a result, a reduction of the variance of the gradients is achieved, which permits to use fewer samples per iteration compared to a standard algorithm, while still preserving the asymptotic convergence rate. Our focus is on the minimization of the Conditional Value-at-Risk (CVaR), and we establish the convergence of the proposed computational framework. Numerical experiments confirm the substantial computational savings achieved by our approach.