Machine learning systems often require updates for various reasons, ranging from the availability of new data or models to the necessity of meeting new or revised technical and ethical metrics. Typically, such metrics capture average performance rather than sample-wise behavior. However, improvements in aggregate measures such as accuracy may lead to negative flips, i.e., instances where the updated model produces errors that the previous model did not, thereby introducing backward incompatibilities. Furthermore, if the distribution of these negative flips is biased with respect to sensitive attributes (e.g., sex or race), the resulting models become not only backward incompatible but also unfair. In this paper, we introduce a generalized notion of Fair Backward Compatibility, which extends and unifies prior work on fairness and backward compatibility. Building on this concept, we propose Fair Backward-Compatible Empirical Risk Minimization, a framework that integrates fairness-aware backward compatibility into most modern machine learning algorithms. We establish the statistical consistency of our framework by showing that both the empirical risk and the empirical Fair Backward Compatibility of the learned model converge to their population-level counterparts at the same rate as in classical Empirical Risk Minimization. We demonstrate that Fair Backward-Compatible Empirical Risk Minimization can be deployed through simple modifications of the cost function in standard machine learning models, using different—convex or at least differentiable—relaxations. We propose a new procedure for tuning and evaluating the performance of models that address both risk and Fair Backward Compatibility. Finally, experiments on real-world datasets, employing both shallow and deep architectures, confirm the effectiveness of our proposal.
Hyper-relational knowledge graph (HRKG), equipped with structures such as superedges, can represent n-ary relations and additional attributes, allowing for a more comprehensive depiction of complex facts and scenarios. Existing methods for constructing KGs are typically optimized for specific dataset, resulting in models with insufficient generalizability. To address this gap, we take full advantage of the multi-task generalization capabilities of large language model (LLM) and propose HKG-MLLM, a framework utilizes multiple lightweight open-source LLMs to collaborate and extract knowledge from text based on highly directive prompts, thereby constructing HRKGs. The framework consists of three key modules: (1) Prompt optimization, which generates optimized prompts based on pre-labeled data. (2) HRKG extraction, which extracts information using the optimized prompts from the previous step. (3) HRKG correction, which ensures the accuracy and consistency of the constructed graph. Our experiments on HyperRED demonstrate that HKG-MLLM achieves significant improvements in precision, recall, and F1 score compared to existing methods based on LLM. This framework effectively addresses the generalizability issue and provides a robust solution for knowledge extraction and representation, paving a way for more versatile and accurate HRKGs.
Accurate voltage estimation is essential for real-time control and condition monitoring of industrial inverter-driven systems, particularly under dynamic operating conditions. This study proposes a sensorless, data-driven framework for estimating phase voltages in a three-phase inverter using advanced machine learning techniques uniquely integrating PSO-driven hyperparameter optimization within a stacking ensemble architecture, an approach not previously explored for inverter voltage estimation. The proposed approach is validated using a comprehensive experimental dataset comprising approximately 235,000 samples, collected under both steady-state and transient conditions. Key input features include DC-link voltage, input voltage, motor speed, phase currents, and Pulse-Width Modulation (PWM) duty ratios. Among individual models, XGBoost demonstrates superior performance with an RMSE of 2.97 V, MAE of 2.05 V, and R2 of 0.999, outperforming Decision Tree, Random Forest, Gradient Boosting, and neural network models. To further enhance prediction accuracy, a stacking-based ensemble framework is developed and optimized using Particle Swarm Optimization (PSO), employing 7 particles over a maximum of 20 iterations with early stopping after 5 non-improving iterations. The final optimized ensemble achieves an improved RMSE of 2.87 V, MAE of 2.04 V, and R2 of 0.999, with prediction errors consistently below 1.6
Few-shot semantic segmentation (FSS) faces challenges due to limited annotated data, leading to performance bottlenecks. However, in existing methods, spatial details suffer from irreversible loss at the early inference stage, forming a bottleneck that hinders performance breakthroughs. This paper introduces NERA-Net, a novel framework designed to address these challenges by focusing on intrinsic feature enhancement. NERA-Net employs a Nested Pyramid Refinement Module (NPRM) to recover spatial details and a Prototype-Pixel Semantic Alignment (PPSA) module to enhance feature discriminability. Additionally, the Generalized Parametric Rectified Linear Unit (GPReLU) is introduced to optimize gradient flow. Experiments on PASCAL-5i and COCO-20i benchmarks demonstrate that NERA-Net achieves state-of-the-art performance, with significant improvements in mIoU (e.g., +2.0
Determining optimal timing for mechanical ventilation weaning remains a persistent clinical challenge, particularly in patients managed for upper airway pathologies—including post-operative head and neck surgery, laryngeal dysfunction, and obstructive airway conditions—where standard weaning criteria derived from general ICU populations may inadequately capture disease-specific physiological dynamics. We present an IoT-driven cyber-physical framework that continuously acquires and fuses four heterogeneous bedside data streams—ventilator waveform parameters, diaphragm ultrasound imaging features, arterial blood gas indices, and patient baseline profiles—within a gated recurrent architecture augmented by cross-modal attention. A Conservative Q-Learning agent learns a dynamic ventilation parameter adjustment policy from retrospective offline data, while a discrete-time survival model provides calibrated, uncertainty-aware individual weaning success probability estimates. To support privacy-preserving multi-center deployment across otolaryngology and head and neck surgery units, the framework employs a FedProx federated learning protocol with Gaussian differential privacy ( ε = 1.0 ) and MAML-style local personalization. Evaluated on 36,181 weaning episodes from MIMIC-IV and eICU-CRD, the system achieves an AUROC of 0.893 on internal validation and 0.871 on external validation, with the federated variant narrowing the gap to centralized training to within 0.001 AUROC. These results demonstrate that integrating IoT-scale multimodal sensing with offline reinforcement learning and federated optimization yields a generalizable decision support system applicable to airway-complex patient populations encountered in otolaryngological practice.
Subgraph extraction constitutes a key component of graph representation learning, enabling scalable analysis and downstream inference over large and complex networks. Nevertheless, existing local extraction approaches frequently exhibit substantial performance degradation in dense graph regimes, where excessive neighborhood expansion and the loss of structural cohesion negatively affect both computational tractability and representational fidelity. Recent work has reformulated the subgraph extraction task in this problem as a local clustering procedure based on a personalized PageRank. However, despite obtaining better results, this approach is affected by the density of the network, limiting its results to networks with medium-low density. In this work, we introduce return random walk kinship (RRWK), a density-aware subgraph extraction framework based on bounded outbound and return random-walk connectivity. The proposed method preserves structurally cohesive regions by exploiting alternative return paths capable of jointly capturing local and global organizational properties of the graph. In contrast to conventional enclosing-subgraph and personalized PageRank-based strategies, RRWK is specifically designed to maintain structural robustness as graph density increases. Furthermore, we study the structural and computational bounds of the proposed method. Experimental results on both synthetic and real-world datasets demonstrate that RRWK consistently preserves the structural properties of the original graph more accurately than state-of-the-art subgraph extraction baselines, while maintaining competitive computational performance in dense graph scenarios.
Detecting small unmanned aerial vehicles (UAVs) in long-range visible-light imagery remains challenging because the targets occupy only a few pixels, have low signal-to-noise ratios, and are strongly affected by cluttered backgrounds and visually similar objects. To address these issues, we propose FLAD, a three-stage fusion-based detector built on a YOLO11-P2 baseline. The design targets a distinct failure mode at each stage of the detector: contour degradation in shallow features, clutter-induced ambiguity in intermediate representations, and localization instability in the detection head. First, LoGStem introduces Laplacian-of-Gaussian (LoG)-initialized edge enhancement together with Gaussian-initialized residual smoothing to strengthen contour cues and suppress background noise in shallow features. Second, RFAConvLSKBlock combines receptive-field attention convolution and large selective kernel (LSK) spatial attention to enhance contextual perception for tiny or weakly textured targets. Third, we adopt LSDC-Head, a lightweight decoupled detection head that uses shared detail-enhanced convolution (DEConv) blocks with group normalization and learnable per-level regression scaling, while retaining the YOLO-style distribution-based bounding-box regression formulation. Extensive experiments on three public benchmarks demonstrate consistent gains over the baseline: mAP@0.5:0.95 improves from 67.3 to 70.4
Dynamic modeling of coal-fired boiler–turbine systems is crucial for flexible and safe operation under high renewable energy penetration. However, strong nonlinearity, multivariable coupling, and variable-specific multi-scale dynamic responses pose significant challenges to accurate prediction. Although data-driven models can achieve high prediction accuracy, their limited transparency restricts their engineering applicability. To address these issues, this study proposes a physics-aligned deep learning framework, termed Multi-dimensional Dynamic Attention Network (MDANet), for boiler–turbine dynamic modeling. In MDANet, adaptive receptive-field temporal feature extraction is designed to characterize variable-specific temporal dynamics, adaptive feature re-weighting is introduced to adjust the relevance of different variables under changing operating conditions, and a temporal context-aware multi-gate mixture-of-experts structure is employed to represent coupled multi-output prediction tasks. The proposed model is validated using DCS data from a 600 MW coal-fired boiler–turbine unit. Comparative experiments, ablation studies, and out-of-sample robustness analysis demonstrate that MDANet achieves accurate and stable prediction performance for unit load, main steam pressure, and main steam temperature. Furthermore, attention-based visualization and quantitative interpretability analysis show that the learned model-internal relevance patterns are consistent with the expected dynamic behavior of the boiler–turbine system. The results indicate that MDANet provides an effective and transparent data-driven modeling approach for the considered boiler–turbine dynamic prediction task.
In modern cybersecurity settings, prioritizing massive streams of incoming network traffic is challenging due to uncertainty and inherent hierarchical dependencies among features and threat indicators. Although intuitionistic fuzzy soft sets (IFSS) offer a powerful framework for handling uncertainty, existing approaches lack effective mechanisms for comparing structured domains while preserving order relations, which may lead to reduced consistency and interpretability of prioritization outcomes. To address this gap, this work presents an order-preserving structural framework for message prioritization in the IFSS paradigm. The study incorporates three key components: (i) a similarity mapping mechanism for comparing ordered IFSS structures with same-level and compatible-level structural comparisons, (ii) the notion of order-isomorphism to capture structural equivalence between domains, and (iii) a topological sorting–based approach to identify the optimal feature–threat correspondences while maintaining hierarchical dependencies. An integrated prioritization pipeline is designed and experimentally validated on the NSL-KDD dataset. Experimental results demonstrate that the proposed approach achieves competitive and balanced performance, attaining the highest test accuracy (0.8365) and F1-score (0.8383), along with improved recall (0.7444), which is important for reducing undetected attacks. The main empirical advantage is reflected in the improved detection of attack instances and the reduced number of false negatives, while statistical significance testing, sensitivity analysis, and perturbation analysis further support the robustness and reliability of the proposed prioritization mechanism. These findings highlight that the integration of order-preserving similarity, structural equivalence, and topological reasoning within IFSS leads to more reliable, consistent, and interpretable prioritization, thereby supporting its relevance in practical applications and decision-making scenarios.
Accurate skin lesion segmentation is essential for early melanoma diagnosis, but remains challenging due to low contrast and ambiguous lesion boundaries. Current deep learning methods often employ standard attention mechanisms with binary gating, which may not fully capture the gradual transitions between lesion and healthy skin. To this end, we propose Fuzzy Attention-aided Encoder-Decoder (FAED) for improved skin lesion segmentation. This architecture integrates soft fuzzy attention mechanism into a U-Net backbone, replacing standard skip connections with a Boundary-conditioned Soft Fuzzy Attention (BSFA) module. BSFA utilizes learnable Gaussian membership functions to model feature relevance probabilistically, rather than using binary selection. Additionally, we introduce an Adaptive Fuzzy Mixture-based aggregation to handle feature variations. Additionally, a Boundary Cue is employed to explicitly refine the attention focus at the lesion perimeters. In addition, empirical profiling on an NVIDIA Tesla T4 shows that FAED performs inference in 10.05 ms/image at batch size 1 and 6.76 ms/image at batch size 8, while requiring 505 MB and 948 MB peak GPU memory during inference, respectively. The proposed architecture maintains low computational cost (2.4M parameters, ∼ 4 GFLOPs) while improving segmentation performance. Evaluations on ISIC2018, ISIC2017, ISIC2016, and PH ^2 datasets demonstrate that FAED achieves competitive performance, yielding Dice scores of 0.9140 on ISIC2016, 0.9135 on PH ^2 , 0.8781 on ISIC2018, and 0.8615 on ISIC17. Ablation studies confirm that both prototype-based aggregation and boundary-conditioned modulation contribute to the observed improvements. These results suggest that modeling feature fusion as soft similarity-based membership, combined with explicit boundary guidance, can improve skin lesion segmentation without increasing architectural complexity.
This paper proposes Deep-Sea Anglerfish Symbiosis Optimization (DASO), a heterogeneous metaheuristic for complex engineering optimization. DASO addresses premature convergence and population diversity loss in high-dimensional, non-convex landscapes by assigning differentiated search roles to male and female agents. Its framework integrates functional dimorphism, asymmetrical attachment, local field-guided movement, and nonlinear high-pressure adaptation to coordinate global exploration and local refinement. DASO was evaluated on the CEC2017 and CEC2022 benchmark suites in 10-, 30-, and 50-dimensional settings, with 30 independent runs for each configuration, and was further tested on constrained engineering design problems. The results show that DASO achieves competitive performance in solution quality, convergence behavior, and robustness compared with representative metaheuristic algorithms. Statistical analysis using the Wilcoxon rank-sum test at the 0.05 significance level further confirms the consistency of the results. These findings demonstrate the potential of heterogeneous role-structured search for solving complex engineering optimization problems.
Chinese Named Entity Recognition (CNER) aims to identify and classify entity mentions from Chinese text and serves as a fundamental task for many downstream natural language processing applications. Recently, word-character lattice frameworks have been widely used to incorporate word-level information into CNER models. However, existing methods typically integrate lexical information through dedicated encoder architectures, such as Flat-Lattice Transformer, while overlooking the complex interactions across different granularities between character-level and word-level representations. Consequently, these methods do not fully exploit lexical knowledge and semantic dependencies, particularly their interactions with character representations. In this paper, we propose A noveL cross-granulariTy contrAstive learnIng network (ALTAI), which integrates lexical and semantic knowledge at multiple granularities. Specifically, ALTAI employs a Cross Transformer to explicitly model interactions among character representations, word-level lexical representations, and semantic representations, enabling effective information exchange across different granularities. Meanwhile, Cross-Granularity Contrastive Learning is introduced to align character representations with their corresponding lexical and semantic views, encouraging consistent representations across character, word, and semantic spaces. Together, these components produce discriminative entity representations by modeling cross-granularity interactions. Experiments on four benchmark datasets show that ALTAI consistently outperforms strong baseline models.
Traditional active noise control research begins by fixing secondary source positions before estimating the secondary path to derive anti-noise signals for primary noise cancellation. However, this method requires the identification of the secondary path, while introducing significantly increased computational complexity due to strong coupling between secondary source paths. To address these limitations, this paper proposes a distributed active noise control optimization method that, for the first time, adopts an optimization-driven perspective to enhance system performance by simultaneously optimizing secondary source placement and initial phases. The framework treats the open-field environment as boundary conditions, while considering secondary source configuration (spatial coordinates and initial phase) as optimization variables to minimize sound pressure levels in target zones. Given the high-dimensional nature of control variables and nonconvex characteristics of the objective function, an Interior Point Method-Assisted Mean Differential Evolution with Weibull distribution (IPMDEW) algorithm is developed to achieve global-local optimization synergy. The hybrid solver integrates a mean differential evolution algorithm with Weibull distribution (MDEW) for global exploration, combined with an interior point method for gradient-based refinement as the population approaches convergence. The results of joint MATLAB-COMSOL simulations verify that the proposed method achieves a superior average noise reduction of 17.2 dB compared to the MDEW algorithm.
Hypertension is a major health challenge that causes premature deaths in the world. Although there has been progress in machine learning and deep learning to predict hypertension, the current approaches still have a number of critical limitations such as extreme class imbalance, suboptimal classification error, high annotation, and poor hyperparameter optimization tools. Most investigations make use of traditional gradient-based or manually-tuned optimization methods, which in most cases lead to local optimal solutions, slow training behavior, inadequate search space exploration, and higher computational costs. These weaknesses associated with optimization reduce the robustness of the models, their stability, and their generalization capabilities in different clinical environments. To mitigate these challenges, an advanced synthetic sampling strategy was employed to address class imbalance by generating representative minority class samples based on distribution-aware weighting mechanisms. This approach reduces bias toward the majority class, improves the structural representation of minority instances within the feature space, and enhances overall classification performance and generalization capability. Building on this balanced dataset, we proposed three novel architectures: Model 1: HighCeptNet, Model 2: HighCeptALNet, and Model 3: FireCeptNet. Model 1 integrates Highway and Inception networks to jointly exploit deep residual learning and multi-scale feature extraction, improving discriminative capacity. Model 2 extends this by incorporating uncertainty-based active learning with Monte Carlo Dropout active learning, enabling selective sampling of the most informative unlabeled instances, which substantially reduces labeling costs while retaining high predictive performance. Model 3 further enhances stability and convergence by integrating firefly optimization for adaptive hyperparameter tuning, thus addressing training instability and improving robustness. Simulation results demonstrate that all three models significantly outperform existing and baseline methods across multiple evaluation metrics. Specifically, Model 1 achieves relative improvements of 3.30
Weakly supervised semantic segmentation (WSSS) aims to produce pixel-level predictions from image-level labels, but it often suffers from incomplete object activation and background ambiguity due to the inherent bias of class activation maps (CAMs). Existing CLIP-based methods improve semantic alignment, they struggle to jointly capture fine grained local details and long-range global dependencies, leading to fragmented activations and blurred boundaries. To address this, we propose a CLIP-based WSSS framework with multi-scale semantic enhancement attention (MSEA) module. MSEA combines multi-scale depthwise separable convolution (DSConv) for local feature extraction, linear attention for efficient global interaction, and an efficient attention (EA) refinement mechanism to suppress noise and enhance boundary quality. And adopt structured attribute embeddings for better semantic guidance. Experiments show that our method outperforms existing single-stage approaches, achieving 75.4
Clinical CT image enhancement requires a precise balance between noise suppression and the preservation of anatomically relevant structures. Conventional genetic algorithms address this as a global optimisation problem, but they typically rely on static rotation schedules that are insensitive to the population’s collective convergence state, leading to premature diversity collapse in high-dimensional parameter spaces. To address this limitation, we propose a Binary Quantum Genetic Algorithm with Diversity-Guided Phase Transitions (BQGA). The algorithm encodes five CT enhancement parameters, including window center, window width, contrast gain, noise suppression, and edge retention, as quantum bits on the Bloch sphere. At each generation, a composite diversity score is computed from four complementary population metrics and mapped to one of three thermodynamic phases (plasma, liquid, crystal) via a hysteresis-buffered threshold rule. The active phase governs the quantum walk step size, mutation rate, and coin operator, providing closed-loop regulation of the balance between exploration and exploitation without manual schedule design. A gradient-aware mutation weight additionally modulates the mutation strength according to the overall edge density of the volume, limiting the introduction of halo artefacts during local refinement. BQGA is evaluated on abdominal and lesion-bearing CT volumes from the CHAOS and LiTS2017 datasets against both metaheuristic optimisers (QGA, QPSO, HHO, MPA) and learning-based methods (Zero-DCE and DPM). It attains the highest mean PSNR, SSIM, and UQI, with the margin widening on low-contrast, high-noise volumes, and non-parametric tests indicate that these gains are statistically significant for most fidelity metrics. Convergence and stability analyses, together with a downstream Chan–Vese segmentation task, further support the reliability and practical utility of the proposed method. Ablation analysis confirms that the phase-transition mechanism accounts for the majority of the performance gain, and that both the multi-dimensional diversity signal and the gradient-aware weighting contribute independently.
Existing personalized federated learning (PFL) methods mostly emphasize accuracy while often overlooking privacy leakage under an honest-but-curious server. Moreover, privacy-preserving PFL faces a practical trade-off between utility and the heavy communication overhead induced by homomorphic encryption. To address these challenges, we propose FedFIHE, which leverages Fisher-information-guided adaptive partitioning to dynamically select shared parameters for global aggregation while retaining the remaining parameters for local personalization under heterogeneous data. To make encrypted training efficient, FedFIHE further introduces a packet-level selective encrypted aggregation scheme. Specifically, clients pack model parameters into packets and upload only the top- κ selected packets with binary masks, and the server performs mask voting to aggregate and broadcast only the most frequently selected packet positions in ciphertext, avoiding the excessive cost caused by inconsistent sparse selections across clients. Extensive experiments on natural image and medical datasets demonstrate that FedFIHE effectively mitigates privacy leakage while achieving better or comparable personalized performance to strong baselines. FedFIHE also remains robust under high sparsity, making it suitable for privacy-sensitive applications such as healthcare.
A fuzzy radial basis inference network with multimodality prior feature knowledge embedding is proposed for multimodality feature fusion and small sample set classification. This model mainly consists of a multi-channel modal feature input layer, a fuzzy radial basis neuron network (FRBN) layer, a modal category feature aggregation layer, a fuzzy rule layer, and a T-S fuzzy classifier. For the differences in spatiotemporal distribution, information granularity, and semantic representation of different modal features, FRBNs is used to embed multimodality diversity prior category feature knowledge, as well as to process the semantic and spatiotemporal information consistency of input modal features. Based on fuzzy computing logic, inference rules are established and multimodality features are fused layer by layer. Specifically, for classification tasks, high-order semantic features of each modality are extracted separately, and a similarity measurement function for modal features is constructed. Then, the fuzzy dynamic C-means clustering algorithm is used to select representative sample features from subsets of different categories in each modality, in order to implicitly represent prior category knowledge. They are used as the kernel center of FRBN to achieve the embedding of diversity prior feature knowledge. In the modality category feature aggregation layer, the diversity feature information of each modality category subclass is aggregated towards the category and generates a non-convex class interface. Fuzzy multiplication operation is adopted to establish fuzzy reasoning rules for classification tasks, and multimodality classification is implemented based on T-S fuzzy classifier. The proposed method can embed and utilize prior feature knowledge of each modality in a mechanism, impose structural and data constraints on the model, reduce the requirement for the completeness of the sample set, and maintain semantic consistency between modalities and within modalities. Applying the proposed method to medical diagnosis, four types of heart disease classification were performed using multimodality medical data such as echocardiogram, multi-lead electrocardiogram, myocardial enzyme examination, and clinical symptoms, with an accuracy rate of 83.37
Multimodal financial signals spanning textual news, market/macro factors, and auxiliary metadata—offer complementary views of investor attention and risk, but exploiting them at scale increasingly relies on large language model (LLM) agents whose multi-step reasoning and verbose generations can be prohibitively costly. We present MENAP (Multimodal Efficiency-constrained News-Augmented Pricing), a cost-aware preference-efficient framework that integrates multimodal inputs into the canonical “news-to-state-to-pricing-to-portfolio” pipeline while explicitly optimizing the news interpretation agent for both signal quality and inference efficiency. MENAP treats each daily multimodal interpretation as a trajectory and performs offline preference optimization, where preferred trajectories are further regularized by efficiency rewards that penalize total token usage and the number of refinement steps, encouraging concise yet informative multimodal summaries. Crucially, the downstream mixed pricing network and evaluation protocol remain unchanged, enabling fair comparisons and straightforward deployment. Experiments follow a standard construction: 2 years of Wall Street Journal news (2021-09-29 to 2023-09-29) aligned with daily returns from CRSP, market and risk-free rates from the Ken French library, and macroeconomic factors following Jensen et al., with a fixed split of 9 months for training, 3 months for validation, and 1 year for testing. We evaluate MENAP on both portfolio performance (Sharpe ratio and maximum drawdown for TP/EW/VW portfolios) and pricing accuracy (average absolute alpha, t-statistics, and the GRS test on 78 anomaly portfolios), and additionally report LLM efficiency metrics (tokens and steps). The results show that MENAP delivers a superior effectiveness–efficiency trade-off, reducing inference cost while maintaining or improving economic outcomes.
The container relocation problem (CRP) is an important combinatorial optimisation problem frequently encountered in container ports. To date, most studies have focused on the static variant of the problem, in which containers are solely retrieved from the yard. In contrast, real-world problems often involve containers arriving and being retrieved simultaneously, a scenario referred to as the dynamic CRP. Solving this dynamic variant using traditional optimisation methods can be challenging, as information about container arrivals is typically not known in advance. As a result, simple constructive heuristics, known as relocation rules (RRs), represent a better alternative for addressing this problem. However, RRs are highly domain specific, making the manual design of effective rules difficult. Therefore, this study employs genetic programming (GP) to generate new RRs tailored to the dynamic CRP automatically. To evaluate their performance, these automatically generated RRs are compared with several manually designed RRs across a wide range of dynamic problem scenarios. The results show that automatically generated RRs significantly outperform their manually designed counterparts, demonstrating the effectiveness of the proposed approach.
