This study introduces the ordered weighted average budget (OWAB) operator, a novel aggregation framework that integrates objective financial data with hierarchical attitudinal preferences in budgeting. Building on classical ordered weighted averaging (OWA) operators and advances in fuzzy logic and generalized aggregation, OWAB provides a flexible mechanism for synthesizing budget components across multiple organizational levels and stakeholder perspectives. The framework encompasses traditional budget aggregation methods as special cases and extends them through generalized and quasi-arithmetic variants, enabling nuanced modeling of optimistic, pessimistic, and neutral attitudes in budget evaluation. Applications to multi-person and multi-country budget aggregation demonstrate the operator’s ability to handle complex, multitiered decision environments. An empirical case study of Australian regional and international research and development (R D) budgets illustrates the practical usefulness of OWAB in supporting transparent, consensus-oriented resource allocation. Taken together, the proposed framework advances both the theoretical foundations and practical methodologies of budget analysis in diverse economic and policy settings.
This paper addresses the pulse-load disturbances introduced by the downstream inverter-side H-bridge. The front-end LLC converter typically suffers from large output-voltage drop, long recovery time, and significant steady-state ripple under such operating conditions. Conventional linear control cannot achieve a satisfactory trade-off between dynamic response and steady-state accuracy. To solve this problem, a switching control strategy based on H-bridge mode identification is proposed. First, a state-plane model of the LLC converter under pulse-load conditions is established to analyze the optimal trajectory radius and switching instants during the transition from no-load to full-load operation. Then, the switching states of the H-bridge are used to construct a load-mode criterion, according to which the controller switches among three operating modes: no-load, full-load, and load-transition modes. Fast regulation is achieved through state-trajectory transition at the instant of load variation. The results show that the proposed strategy effectively suppresses voltage overshoot and drop without requiring additional hardware. The dynamic settling time is reduced by 50
Fuzzy multi-criteria decision-making (MCDM) methods such as AHP and TOPSIS are widely used to evaluate complex alternatives under uncertainty. The introduction of Z-numbers, incorporating both fuzzy restriction and reliability information, has further extended the expressive capability of these models. However, despite their growing application, computational realisations of Z-number fuzzy AHP–TOPSIS are often presented as procedural workflows in which modelling assumptions remain implicit and difficult to isolate. This paper presents an R-based computational implementation and framework prototype that formalises established triangular Z-number AHP–TOPSIS procedures within an explicit, script-controlled architecture. The framework preserves the conventional mathematical structure while exposing key modelling components - linguistic scale specification, reduction of Z-numbers, weight defuzzification, and ranking configuration - as selectable parameters. The approach is demonstrated through the computational replication of a published Z-number fuzzy AHP–TOPSIS study, with all data and code provided in an online repository to enable independent execution of the complete decision pipeline. By separating representation, transformation, aggregation, and ranking stages, the framework improves transparency and analytical traceability without altering the theoretical foundations of Z-number fuzzy decision-making.
Vision-Language Models (VLMs) show great potential in autonomous driving. However, they often operate as black boxes and fail in safety-critical scenarios. Physical hallucinations and causal confusion cause these failures. To address these Explainable AI challenges, we propose PG-Agent (Physics-Grounded Agent). This framework enhances robust decision making under uncertainty. It uses a neuro-symbolic approach to align neural perception with symbolic causality. Traditional methods rely on implicit feature matching. Our approach explicitly grounds the VLM through a dual-stream mechanism. First, we employ WideGazeNet. This lightweight network generates semantic-aware visual prompts to simulate human visual attention and filter environmental noise. Second, we align this visual stream with a Structural Causal Model derived from simulator ground truth. This step translates expert physical laws into Chain-of-Thought supervision. We also introduce a visual counterfactual training strategy. It physically removes risk objects from the scene. This intervention forces the agent to learn the true causal link between object existence and actions. Experiments on the DATAD benchmark show that PG-Agent achieves top performance in lateral control. It demonstrates highly interpretable and real-time decision making in complex scenarios.
Addressing the critical challenge of detecting incipient bearing faults in low signal-to-noise ratio (SNR) environments, a Health Indicator (HI) framework is formulated utilizing Crest Factor-driven chaotic synchronization. By exploiting the inherent sensitivity of chaotic systems to infinitesimal perturbations, the methodology nonlinearly amplifies weak, fault-related impulses typically submerged in heavy background noise. The Crest Factor (CF) is integrated as a dynamic driver for a Lorenz chaotic system, where the resulting synchronization error energy serves as a high-sensitivity HI for degradation assessment. Empirical validation on the XJTU-SY dataset yields an Early Detection Percentage (EDP) of 18.6
Language models (LLMs) are promising for recommender systems due to their parametric knowledge and natural language reasoning over user preferences. However, direct LLM prompting incurs high inference costs and context collapse, where long user histories inflate prompts—driving latency and hallucinations—while yielding shallow, pattern-matching recommendations. We introduce a multi-agent memory framework that distills long and short-term user interactions plus item audience signals into dual-format profiles: structured features paired with natural-language summaries. A Mamdani fuzzy inference engine aggregates signals across memory dimensions to derive compatibility scores, constraining LLM context and offloading scoring. This design further enables incremental real-time updates to user/item representations, obviating full-dataset retraining. Crucially, how fuzzy scores are presented to the LLM matters as much as the scores themselves: a fuzzy-first prompt strategy—where the LLM defaults to the top-ranked candidate and overrides only with concrete evidence—outperforms both standalone fuzzy scoring and naive score injection. On MovieLens 1M, the combined system attains 60.5 × over direct prompting—while reducing token consumption ∼80% and inference time ∼3.6× across LLM backbones.
This paper proposes an Evidence-Driven Game-Theoretic Decision-Making Consensus Model (EDCB) to address the practical challenges in group decision-making where significant differences in experts’ subjective preferences hinder stable consensus formation, and traditional consensus methods lack objective guidance mechanisms, often triggering opinion conflicts and complicating the implementation of decisions. First, the Analytic Hierarchy Process (AHP) is employed to map the structure of expert subjective preferences. Then, it constructs an adaptive step function based on subjective-objective bias to precisely correct deviations, driving iterative convergence of expert preferences toward the data-based benchmark through multiple rounds of iteration. Subsequently, a behavioral consensus criterion rooted in game steady states is proposed, incorporating individual utility functions and adopting Nash equilibrium as the stopping rule for iterations. Finally, the TOPSIS method is applied to select the optimal alternative from the consensus weights after meeting standards. Experimental results validate the effectiveness of the EDCB and demonstrate its effectiveness through a case study on evaluating marketing strategies for retail enterprises.
A secondary parallel CLLC resonant converter is proposed to address the issue of traditional CLLC resonant converters lacking current source output characteristics. The proposed converter has two working modes of constant current and constant voltage, making it highly suitable for applications such as battery charging. In constant current mode, the proposed converter adjusts the output current by changing the phase shift angle of the full bridge inverter; In constant voltage mode, the voltage gain of the resonant network is adjusted by changing the switching frequency of the full bridge inverter. Feedforward fuzzy control is proposed to address the drawbacks of PID control, such as difficulty in parameter tuning without a model, mediocre control effect on strongly nonlinear systems, and sensitivity to noise. The proposed control strategy has the advantages of rapid response, small overshoot, fast convergence, and strong robustness. To verify the feasibility of the proposed converter and control strategy, a simulation model was built using PLECS software for validation. The results showed that the relevant characteristics of the proposed converter and control strategy were in line with theoretical analysis.
This paper proposes a reachable-set-based receding-horizon trajectory planning method for unmanned surface vessels (USVs). A dynamic model is established to capture the coupled translational and rotational motions of the USV. Based on this model, forward reachable sets over a finite prediction horizon are constructed to characterize all dynamically feasible future states. A rolling planning framework is then implemented, wherein an optimal state is selected at each planning step by solving a quadratic programming (QP) problem. The designed cost function jointly accounts for position convergence toward the target, heading alignment, and angular velocity smoothness, thereby ensuring both navigation accuracy and maneuvering stability. To further enhance robustness near the target region, a backtracking strategy is incorporated to prevent oscillations and performance degradation during the final approach. Extensive Simulation results under different initial conditions and environmental disturbances demonstrate that the proposed method generates smooth, dynamically feasible trajectories and exhibits strong adaptability and robustness in complex scenarios.
Many real-world decision systems require staged, knowledge-guided sequential consultation of information sources—e.g., KYC verification, medical diagnosis with progressive testing, or fraud detection with escalating investigation depth. Standard decision trees freely mix features at each depth level, preventing efficient implementation of such staged systems and making them incompatible with the knowledge engineering principle of structured, interpretable decision hierarchies. We propose Restricted Decision Trees (RDT), which enforce feature consistency at each depth level while allowing node-specific threshold adaptation—a structure that directly encodes the domain knowledge of “which information source to consult at each stage.” Comprehensive evaluation across 12 diverse datasets shows RDT achieves 99.1
In group decision-making (GDM), expert opinions are increasingly represented as probability distributions to capture inherent uncertainty. Existing dominant aggregation approaches, including linear combination and quantile averaging, are predominantly additive and rely on exogenously specified weights, limiting their ability to reflect interaction patterns among experts. Moreover, although trust has been widely studied in social network–based GDM, its integration into distribution-based aggregation remains underexplored. To address these issues, this paper proposes a consensus–trust-driven probabilistic aggregation framework. First, distributional conflict between experts is quantified via the 1-Wasserstein distance in the quantile domain, and a conflict-sensitive trust function is constructed to capture relational credibility decay. Second, consensus and trust are fused into cooperation intensities, which are embedded into a 2-additive capacity through normalized Möbius coefficients. Finally, a Choquet-based quantile averaging operator is developed to obtain the aggregated collective distribution while preserving distributional validity and modeling interaction effects. An automotive complaint severity assessment example involving experts and large language models illustrates its effectiveness. The proposed model bridges probabilistic uncertainty modeling and interaction-aware aggregation theory, offering a structurally interpretable and flexible framework for complex decision scenarios.
With the increasing number of levels in cascaded H-bridge multilevel converter (CHBMC), traditional space vector pulse width modulation (SVPWM) faces severe challenges, such as the enormous computational burden of reference vector location and the accompanying high-frequency common-mode voltage (CMV). This paper proposes a simplified SVPWM and CMV suppression strategy based on a J-K coordinate system for three-phase CHBMCs. First, a fast location algorithm based on J and K parameters is proposed in the Cartesian coordinate system. By mapping the space vectors onto integer coordinates, the underlying computational logic is significantly simplified. Second, through the algebraic analysis of redundant switching states, the optimal switching states with the minimum absolute CMV value are calculated, which effectively reduces the CMV. Simulations based on a 6-cell 13-level CHBMC fully verify the correctness and effectiveness of the proposed strategy, demonstrating that this method achieves excellent CMV suppression performance while remarkably reducing the consumption of computational resources.
The exponential quasi-projective synchronization (QPJS) is investigated about quaternion-valued non-identical Cohen-Grossberg neural networks (QVNNs) featuring interaction terms and mismatched parameters. By constructing appropriate Lyapunov functions and designing an effective nonlinear controller, sufficient conditions for achieving QPJS of the studied QVNNs have been rigorously established. The non-decomposition technique is employed, since separating QVNNs demands sophisticated calculations and derivation steps. Furthermore, the error bound of the considered system is estimated. Finally, numerical examples are simulated to verify the effectiveness and feasibility of the proposed results.
Uncertainty and observation errors are commonly involved in real-valued time series, which may weaken the performance of conventional time series classification methods based only on point-valued observations. To address this issue, this paper proposes an interval gray transformation framework for time series classification. In the proposed framework, each original point-valued time series is first transformed into a stationary series, and the residual information obtained from an ARIMA model is then used to construct interval gray numbers with known bounds. In this way, the original time series is represented not only by its central values but also by its uncertainty range. Based on this representation, the classical DTW-1NN classifier is extended to the interval gray environment, leading to the proposed IGN-DTW-1NN method. Experiments on several datasets from the UCR time series classification archive demonstrate that the proposed method can effectively reduce classification error rates compared with the conventional DTW-1NN method and several representative classification approaches. The results indicate that incorporating residual-based interval gray information provides a useful and interpretable way to improve point-valued time series classification.
Swarm-based algorithms are powerful in solving diverse optimisation problems. However it is still a challenging task for any single metaheuristic optimisation algorithm to perform optimally across all benchmark landscapes. As an example, while Particle Swarm Optimisation (PSO) often converges quickly on smooth unimodal functions, it may suffer from premature convergence on multimodal landscapes with many deceptive local optima. Therefore, this research aims to propose an adaptive hybrid optimisation algorithm in which a Q-Learning method selects complementary optimisation algorithms. Specifically, the proposed model uses a PSO warm-start phase followed by a tabular Q-Learning controller that dynamically selects between Firefly (FA) and Genetic Algorithm (GA). Algorithm selection is determined by population-level characteristics such as normalised diversity and normalised fitness spread. This RL-based search action controller is trained offline to learn an interpretable algorithm-selection policy and is applied with a dynamic Q-table during evaluation. Evaluated using various continuous and complex benchmark functions over multiple independent runs, the proposed algorithm outperforms standalone PSO, GA, FA and Simulated Annealing across diverse landscapes. Statistical validation using the Wilcoxon rank-sum test confirms significant improvements in most comparisons.
In this paper, a learning-based obstacle avoidance problem for unmanned aerial vehicles (UAVs) under perceptual uncertainty is studied, particularly when range-like observations and velocity-related cues are jointly corrupted. In such regimes, nominal safety potentials can become over-optimistic near obstacles, producing dense feedback that underestimates risk and weakens corrective signals from sparse failures. To address this, a plug-in potential-based reward shaping module is proposed. It constructs a bounded safety potential from aligned distance and closing-speed cues, evaluates it on a fixed perturbation set, and aggregates the resulting scores via an ordered Choquet integral induced by a λ -fuzzy measure. This design captures non-additive interactions between coupled perturbations and yields interpretable Choquet weights. To avoid using λ as an ad-hoc conservatism knob, conservatism is defined operationally and the fuzzy-measure parameters are calibrated offline with a near-obstacle objective that penalizes overly high potentials. In pilot-scale simulations on SimpleAvoid across multiple uncertainty levels, the proposed shaping was associated with higher task completion under non-zero noise while maintaining comparable collision rates.
This paper introduces FuzzyRules, an R package for the systematic development, analysis, and evaluation of fuzzy rule-based systems. The framework provides a unified and extensible architecture that supports the full modeling pipeline of interpretable fuzzy inference systems, with emphasis on Artificial Intelligence and classification tasks. The package implements a comprehensive set of tools for the creation and manipulation of fuzzy sets, including set-theoretic operations based on t-norms and dual t-conorms (intersection, union, and complement), as well as structural descriptors such as support, kernel, core, height, degree of membership, overlap functions and indices, grouping functions and indices, and the Type-1 Jaccard fuzzy similarity index. It incorporates a broad collection of membership functions and classical t-norms and t-conorms. For inference, the framework implements Mamdani-type fuzzy systems with multiple defuzzification strategies, including the classical center of gravity, and nine other methods. Methodological foundations and an illustrative example are provided to facilitate reproducible experimentation and practical adoption. A brief comparison with other software is also presented.
Aggregation methods for Gross Domestic Product (GDP) often rely on risk-neutral assumptions and therefore may fail to capture asymmetric policy concerns, where decision-makers may weight downside risks and upside potential differently. This paper applies the Ordered Weighted Averaging (OWA) operator originally introduced by Yager (1988) to the analysis of GDP and introduces the Ordered Weighted Average Gross Domestic Product (OWAGDP) as a preference-sensitive aggregation framework. The approach generates a family of GDP aggregates ranging from worst-case (extreme pessimism) to best-case (extreme optimism), with the arithmetic mean emerging as a special case corresponding to attitudinal neutrality. The paper further extends the framework to generalized and quasi-arithmetic forms, as well as to the multi-person OWAGDP operator, in order to aggregate heterogeneous GDP estimates from multiple experts in non-linear environments. An illustrative example using the European Union GDP per inhabitant data shows that different attitudinal weights produce distinct GDP distributions. This highlights how the OWAGDP framework provides a more flexible, preference-sensitive aggregation approach for decision-making under uncertainty.
User performance assessment is a fundamental component of virtual reality (VR) training systems, particularly in health-related simulators where objective and real-time skill evaluation is required. Single User Assessment Systems (SUAS) rely on probabilistic inference models to estimate user proficiency from interaction data, with temporal variables playing a central role in the assessment process. Since time-based metrics in VR environments often exhibit asymmetric and heavy-tailed behavior, selecting an appropriate statistical distribution is critical for reliable classification. This paper proposes a Fuzzy Weibull Naive Bayes Network (FWNB) as the core inference model of a SUAS, combining the flexibility of the Weibull distribution for modeling continuous time-to-event variables with fuzzy membership functions to handle uncertainty and variability in performance data. The model was evaluated through one thousand Monte Carlo simulation replicas, considering three performance classes, and compared with traditional Naive Bayes (NB) and Weibull Naive Bayes (WNB) approaches. The proposed SUAS achieved 91.11
Although the participation and activity levels of female scientists in academia have increased markedly, existing studies have largely attributed this change to external institutional and cultural drivers, portraying women as passive beneficiaries while overlooking their agentic roles and substantive contributions in knowledge-creation practices. To address this gap, this study introduces the concept of Knowledge-Related Roles (KRRs) and establishes a role classification framework comprising three major categories and nine subcategories, aiming to examine the evolution of female scientists’ roles and the underlying nature of their rising status from the perspective of knowledge production. The study develops a role identification system based on large language model (LLM) driven text mining to systematically extract and quantify the specific roles performed by 1,190 outstanding female scientists, along with their corresponding textual and lexical evidence. By tracing the dynamic changes of these roles across historical periods, the study reveals that the core driver of female scientists’ status advancement lies in the gradual transformation of their knowledge-related roles from “knowledge assistants” to “knowledge leaders”. This finding demonstrates both consistency and meaningful heterogeneity across different disciplines and regional-cultural contexts. Theoretically, this research offers a knowledge-centered perspective for understanding changes in female’s status in science, and it provides profound and insightful implications for the cultivation and development of female talent.
