Applied Soft Computing最新文献_第4页

CSME-net: A lightweight unsupervised low-light image enhancement network based on luminance–chrominance decoupling and cascaded multi-stage optimization CSME-net：基于亮度-色度解耦和级联多阶段优化的轻量级无监督微光图像增强网络

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-23 DOI: 10.1016/j.asoc.2026.114676

Chao Sun , Junqi Xia , Liheng Xia , Jiuye Shi , Jianjun Ding

Low-light image enhancement (LLIE) is a critical technology for ensuring the robustness of downstream vision tasks such as autonomous driving and intelligent surveillance. To address the limitations of existing enhancement methods, including heavy reliance on paired supervised data, high computational cost, and susceptibility to color distortion, we propose CSME-Net, a lightweight, unsupervised, and color-adaptive enhancement network. Operating in the YUV color space, the network adopts a “luminance-prior, chrominance-controlled” channel separation strategy, performing structure-aware adaptive enhancement exclusively on the luminance (Y) channel while applying constrained fine-tuning to the chrominance (UV) channels, effectively reducing redundant computation while maintaining color consistency. Furthermore, we introduce the Y-attention module, which leverages the luminance variation (

Δ

Y) as an online guidance signal to dynamically regulate chrominance enhancement, achieving real-time interaction and balance between luminance and chrominance. By integrating BilateralFReLU activation and Gamma-weighted loss, CSME-Net significantly enhances feature extraction capabilities while maintaining an extremely lightweight architecture with only 0.00495M parameters. Experimental results demonstrate that, under an unsupervised framework, CSME-Net can achieve stable convergence with only 100 training samples. The method not only effectively suppresses color distortion and information loss but also provides a valuable pathway for high-quality, high-frame-rate, real-time image enhancement on low-computation platforms.

低光图像增强（LLIE）是确保自动驾驶和智能监控等下游视觉任务鲁棒性的关键技术。为了解决现有增强方法的局限性，包括严重依赖成对监督数据、计算成本高以及易受颜色失真的影响，我们提出了CSME-Net，一种轻量级、无监督和颜色自适应增强网络。在YUV色彩空间中，该网络采用“亮度优先，色度控制”的通道分离策略，只对亮度(Y)通道进行结构感知的自适应增强，对色度（UV）通道进行约束微调，在保持色彩一致性的同时有效减少冗余计算。此外，我们还引入了y -注意力模块，该模块利用亮度变化（ΔY）作为在线引导信号来动态调节亮度增强，实现亮度和亮度之间的实时交互和平衡。通过整合BilateralFReLU激活和伽玛加权损失，CSME-Net显著增强了特征提取能力，同时保持了一个非常轻量级的架构，只有0.00495M个参数。实验结果表明，在无监督框架下，CSME-Net只需要100个训练样本就可以实现稳定的收敛。该方法不仅有效地抑制了图像的颜色失真和信息丢失，而且为在低计算平台上实现高质量、高帧率、实时的图像增强提供了一条有价值的途径。

{"title":"CSME-net: A lightweight unsupervised low-light image enhancement network based on luminance–chrominance decoupling and cascaded multi-stage optimization","authors":"Chao Sun , Junqi Xia , Liheng Xia , Jiuye Shi , Jianjun Ding","doi":"10.1016/j.asoc.2026.114676","DOIUrl":"10.1016/j.asoc.2026.114676","url":null,"abstract":"<div><div>Low-light image enhancement (LLIE) is a critical technology for ensuring the robustness of downstream vision tasks such as autonomous driving and intelligent surveillance. To address the limitations of existing enhancement methods, including heavy reliance on paired supervised data, high computational cost, and susceptibility to color distortion, we propose CSME-Net, a lightweight, unsupervised, and color-adaptive enhancement network. Operating in the YUV color space, the network adopts a “luminance-prior, chrominance-controlled” channel separation strategy, performing structure-aware adaptive enhancement exclusively on the luminance (Y) channel while applying constrained fine-tuning to the chrominance (UV) channels, effectively reducing redundant computation while maintaining color consistency. Furthermore, we introduce the Y-attention module, which leverages the luminance variation (<span><math><mi>Δ</mi></math></span>Y) as an online guidance signal to dynamically regulate chrominance enhancement, achieving real-time interaction and balance between luminance and chrominance. By integrating BilateralFReLU activation and Gamma-weighted loss, CSME-Net significantly enhances feature extraction capabilities while maintaining an extremely lightweight architecture with only 0.00495M parameters. Experimental results demonstrate that, under an unsupervised framework, CSME-Net can achieve stable convergence with only 100 training samples. The method not only effectively suppresses color distortion and information loss but also provides a valuable pathway for high-quality, high-frame-rate, real-time image enhancement on low-computation platforms.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114676"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080193","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Video segment localization network based on mutual information dual-graph contrastive learning 基于互信息双图对比学习的视频片段定位网络

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-23 DOI: 10.1016/j.asoc.2026.114699

GuangLi Wu , Yulong Cui , Jing Zhang

The task of video segment localization based on audio queries aims to identify video segments that semantically align with a given audio query. While existing methods effectively capture both global and local features, they often overlook the complementary correlations between audio and visual modalities. In this paper, we focus on the semantic consistency and structural divergence between audio-visual modalities and propose a video segment localization network based on mutual information dual-graph contrastive learning. Our method achieves synergistic optimization of consistency modeling and redundancy suppression for graph representations by constructing a perturbed graph view and adopting a pairwise mutual information optimization strategy that maximizes semantic consistency and minimizes redundant information, which significantly improves the discriminative and generalization capabilities of graph-structured representations. To address the challenges of event asynchrony and semantic redundancy, we introduce a Cross-Modal Gated Attention (CMGA) module, which uses audio cues to semantically weight and filter visual features along channel and spatial dimensions, thus enhancing modality alignment and suppressing background noise. Furthermore, a Cross-Gated Attention (CGA) module is proposed to enable bidirectional guidance and dynamic fusion between modalities, strengthening fine-grained cross-modal semantic interactions in terms of robustness and generalization. In addition, we propose a Cross-Gated Attention (CGA) module that integrates cross-attention and dynamic gating mechanisms to achieve bidirectional guidance and adaptive fusion between modalities. By dynamically adjusting the contribution of each modality during feature interaction, thereby demonstrating excellent robustness and generalization performance. Experiments conducted on the widely used AVE dataset demonstrate that our method achieves an accuracy of 77.8%, outperforming mainstream baselines. To further evaluate localization performance and cross-scene generalization under untrimmed video settings, we introduce the UnAV-100 dataset, where our method attains an accuracy of 57.6%, validating its effectiveness and generalization capability.

基于音频查询的视频片段定位任务旨在识别与给定音频查询在语义上一致的视频片段。虽然现有的方法可以有效地捕获全局和局部特征，但它们往往忽略了音频和视觉模式之间的互补相关性。本文针对视听模态之间的语义一致性和结构差异，提出了一种基于互信息双图对比学习的视频片段定位网络。该方法通过构造摄动图视图，采用语义一致性最大化和冗余信息最小化的两两互信息优化策略，实现了图表示一致性建模和冗余抑制的协同优化，显著提高了图结构表示的判别能力和泛化能力。为了解决事件异步和语义冗余的挑战，我们引入了一个跨模态门控注意（CMGA）模块，该模块使用音频线索沿通道和空间维度对视觉特征进行语义加权和过滤，从而增强模态对齐并抑制背景噪声。此外，提出了一种交叉门控注意（Cross-Gated Attention， CGA）模块，实现模态之间的双向引导和动态融合，增强了细粒度跨模态语义交互的鲁棒性和泛化性。此外，我们提出了一个交叉门控注意（CGA）模块，该模块集成了交叉注意和动态门控机制，实现了模态之间的双向引导和自适应融合。通过在特征交互过程中动态调整各模态的贡献，从而表现出优异的鲁棒性和泛化性能。在广泛使用的AVE数据集上进行的实验表明，我们的方法达到了77.8%的准确率，优于主流基线。为了进一步评估未修剪视频设置下的定位性能和跨场景泛化，我们引入了UnAV-100数据集，其中我们的方法达到了57.6%的准确率，验证了其有效性和泛化能力。

{"title":"Video segment localization network based on mutual information dual-graph contrastive learning","authors":"GuangLi Wu , Yulong Cui , Jing Zhang","doi":"10.1016/j.asoc.2026.114699","DOIUrl":"10.1016/j.asoc.2026.114699","url":null,"abstract":"<div><div>The task of video segment localization based on audio queries aims to identify video segments that semantically align with a given audio query. While existing methods effectively capture both global and local features, they often overlook the complementary correlations between audio and visual modalities. In this paper, we focus on the semantic consistency and structural divergence between audio-visual modalities and propose a video segment localization network based on mutual information dual-graph contrastive learning. Our method achieves synergistic optimization of consistency modeling and redundancy suppression for graph representations by constructing a perturbed graph view and adopting a pairwise mutual information optimization strategy that maximizes semantic consistency and minimizes redundant information, which significantly improves the discriminative and generalization capabilities of graph-structured representations. To address the challenges of event asynchrony and semantic redundancy, we introduce a Cross-Modal Gated Attention (CMGA) module, which uses audio cues to semantically weight and filter visual features along channel and spatial dimensions, thus enhancing modality alignment and suppressing background noise. Furthermore, a Cross-Gated Attention (CGA) module is proposed to enable bidirectional guidance and dynamic fusion between modalities, strengthening fine-grained cross-modal semantic interactions in terms of robustness and generalization. In addition, we propose a Cross-Gated Attention (CGA) module that integrates cross-attention and dynamic gating mechanisms to achieve bidirectional guidance and adaptive fusion between modalities. By dynamically adjusting the contribution of each modality during feature interaction, thereby demonstrating excellent robustness and generalization performance. Experiments conducted on the widely used AVE dataset demonstrate that our method achieves an accuracy of 77.8%, outperforming mainstream baselines. To further evaluate localization performance and cross-scene generalization under untrimmed video settings, we introduce the UnAV-100 dataset, where our method attains an accuracy of 57.6%, validating its effectiveness and generalization capability.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114699"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080295","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

DualBranch-YDNet: A dual-stream deep segmentation framework for accurate rice chalkiness detection DualBranch-YDNet：用于水稻白垩度精确检测的双流深度分割框架

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-22 DOI: 10.1016/j.asoc.2026.114690

Jinfeng Zhao , Sheng Dai , Xianglong Xia , Guochao Zhao , Jie Qiu , Xin Wei , Xuehui Huang , Yan Ma , Jie Yang

Accurate segmentation of chalky regions remains a critical and challenging problem in automated rice quality assessment, due to their low contrast, small size, and complex morphology. Existing methods often rely on grain-level classification without pixel-level localization, limiting their applicability in fine-grained phenotyping. We propose DualBranch-YDNet, a novel dual-branch deep segmentation framework that integrates YOLO-SAM for rapid grain localization and DeepLabv3+ for fine-grained chalkiness segmentation. A lightweight mask fusion strategy aligns outputs from both branches, enabling precise delineation of grain contours and internal defects. To address small-object and class imbalance challenges, we incorporate Reparameterized Large Kernel Convolution (RepLKConv) and a composite WBCE–Dice loss, significantly enhancing accuracy and convergence. We also construct a large-scale, high-resolution rice chalkiness dataset covering diverse varieties and spatial arrangements. Experiments demonstrate that DualBranch-YDNet achieves an average error of 1.0% in chalkiness degree and a maximum deviation of 2.0% in chalky grain rate, surpassing four state-of-the-art baselines. These results confirm its robustness, accuracy, and suitability for laboratory-based and off-line high-throughput rice quality evaluation in real-world phenotyping scenarios.

由于白垩质区域对比度低、面积小、形态复杂，对其进行准确分割一直是大米质量自动化评价中的一个关键问题。现有的方法往往依赖于颗粒级分类，而没有像素级定位，限制了它们在细粒度表型分析中的适用性。我们提出了一种新的双分支深度分割框架DualBranch-YDNet，该框架集成了用于快速颗粒定位的YOLO-SAM和用于细粒度白垩度分割的DeepLabv3+。轻量级掩模融合策略将两个分支的输出对齐，从而能够精确描绘颗粒轮廓和内部缺陷。为了解决小对象和类不平衡的挑战，我们结合了重参数化大核卷积（RepLKConv）和复合WBCE-Dice损失，显著提高了精度和收敛性。我们还构建了一个覆盖不同品种和空间排列的大规模、高分辨率水稻白垩度数据集。实验表明，DualBranch-YDNet在白垩度上的平均误差为1.0%，在白垩粒率上的最大偏差为2.0%，优于4条最先进的基线。这些结果证实了该方法的稳健性、准确性和适用性，可用于基于实验室和离线的高通量水稻品质评估。

{"title":"DualBranch-YDNet: A dual-stream deep segmentation framework for accurate rice chalkiness detection","authors":"Jinfeng Zhao , Sheng Dai , Xianglong Xia , Guochao Zhao , Jie Qiu , Xin Wei , Xuehui Huang , Yan Ma , Jie Yang","doi":"10.1016/j.asoc.2026.114690","DOIUrl":"10.1016/j.asoc.2026.114690","url":null,"abstract":"<div><div>Accurate segmentation of chalky regions remains a critical and challenging problem in automated rice quality assessment, due to their low contrast, small size, and complex morphology. Existing methods often rely on grain-level classification without pixel-level localization, limiting their applicability in fine-grained phenotyping. We propose DualBranch-YDNet, a novel dual-branch deep segmentation framework that integrates YOLO-SAM for rapid grain localization and DeepLabv3+ for fine-grained chalkiness segmentation. A lightweight mask fusion strategy aligns outputs from both branches, enabling precise delineation of grain contours and internal defects. To address small-object and class imbalance challenges, we incorporate Reparameterized Large Kernel Convolution (RepLKConv) and a composite WBCE–Dice loss, significantly enhancing accuracy and convergence. We also construct a large-scale, high-resolution rice chalkiness dataset covering diverse varieties and spatial arrangements. Experiments demonstrate that DualBranch-YDNet achieves an average error of 1.0% in chalkiness degree and a maximum deviation of 2.0% in chalky grain rate, surpassing four state-of-the-art baselines. These results confirm its robustness, accuracy, and suitability for laboratory-based and off-line high-throughput rice quality evaluation in real-world phenotyping scenarios.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114690"},"PeriodicalIF":6.6,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080103","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Bi-directional distillation enabling progressive lightweight model training in edge computing-based federated learning 双向蒸馏，在基于边缘计算的联邦学习中实现渐进式轻量级模型训练

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-22 DOI: 10.1016/j.asoc.2026.114712

Yishan Chen , Wenshuo Dai , Zhen Qin , Junxiao Han

Wireless Federated Learning (FL) has emerged as a promising paradigm for enabling collaborative intelligence in edge computing while preserving data privacy by exchanging model parameters instead of raw data. However, deploying large-scale models remains challenging due to the limited computational resources and constrained communication bandwidth of edge devices. To address this issue, the paper proposes a Bi-Directional Distillation-Driven Hierarchical Framework (BDH-LM) that enables Progressive Lightweight Model Training across edge and cloud. Central to our framework is the Streamline Teacher Network (STN) algorithm, which progressively compresses the network size by generating a series of compact student models that replace large-scale models without significant performance degradation. A bi-directional knowledge distillation mechanism is established to facilitate mutual knowledge exchange between the cloud and edge models, thereby reducing the computational burden on cloud servers while preserving user privacy on edge devices. To further minimize communication overhead, a data-value-aware selection strategy is introduced to identify and transfer only the most informative samples. Moreover, we formulate a joint optimization problem that considers energy and latency constraints, optimizing pruning ratios, CPU frequencies, uplink power, and bandwidth allocation to accelerate the training process under resource limitations. Finally, experiments conducted on CIFAR-10, CIFAR-100, and SVHN datasets further validate the theoretical results.

无线联邦学习（FL）已经成为一种很有前途的范例，可以在边缘计算中实现协作智能，同时通过交换模型参数而不是原始数据来保护数据隐私。然而，由于边缘设备的计算资源有限和通信带宽受限，部署大规模模型仍然具有挑战性。为了解决这个问题，本文提出了一个双向蒸馏驱动的分层框架（BDH-LM），该框架支持跨边缘和云的渐进式轻量级模型训练。我们的框架的核心是精简教师网络（STN）算法，该算法通过生成一系列紧凑的学生模型来逐步压缩网络大小，这些模型可以取代大规模模型，而不会显著降低性能。建立双向知识蒸馏机制，促进云和边缘模型之间的相互知识交换，从而减少云服务器的计算负担，同时保护边缘设备上的用户隐私。为了进一步减少通信开销，引入了数据值感知选择策略，以识别和传输信息量最大的样本。此外，我们制定了一个考虑能量和延迟约束的联合优化问题，优化剪枝比、CPU频率、上行功率和带宽分配，以加速资源限制下的训练过程。最后，在CIFAR-10、CIFAR-100和SVHN数据集上进行的实验进一步验证了理论结果。

{"title":"Bi-directional distillation enabling progressive lightweight model training in edge computing-based federated learning","authors":"Yishan Chen , Wenshuo Dai , Zhen Qin , Junxiao Han","doi":"10.1016/j.asoc.2026.114712","DOIUrl":"10.1016/j.asoc.2026.114712","url":null,"abstract":"<div><div>Wireless Federated Learning (FL) has emerged as a promising paradigm for enabling collaborative intelligence in edge computing while preserving data privacy by exchanging model parameters instead of raw data. However, deploying large-scale models remains challenging due to the limited computational resources and constrained communication bandwidth of edge devices. To address this issue, the paper proposes a Bi-Directional Distillation-Driven Hierarchical Framework (BDH-LM) that enables Progressive Lightweight Model Training across edge and cloud. Central to our framework is the Streamline Teacher Network (STN) algorithm, which progressively compresses the network size by generating a series of compact student models that replace large-scale models without significant performance degradation. A bi-directional knowledge distillation mechanism is established to facilitate mutual knowledge exchange between the cloud and edge models, thereby reducing the computational burden on cloud servers while preserving user privacy on edge devices. To further minimize communication overhead, a data-value-aware selection strategy is introduced to identify and transfer only the most informative samples. Moreover, we formulate a joint optimization problem that considers energy and latency constraints, optimizing pruning ratios, CPU frequencies, uplink power, and bandwidth allocation to accelerate the training process under resource limitations. Finally, experiments conducted on CIFAR-10, CIFAR-100, and SVHN datasets further validate the theoretical results.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114712"},"PeriodicalIF":6.6,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080102","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Novel multi-cube extreme learning machine with sparse attention and ridge regression technique for robust machine learning 基于稀疏注意和脊回归的鲁棒机器学习多立方极值学习机

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-22 DOI: 10.1016/j.asoc.2026.114704

Merve Zirekgur , Barış Karakaya , Abdulkadir Sengur , U. Rajendra Acharya

This study introduces a learning framework designed to enhance the representational capacity and stability of single-layer feed-forward networks (SLFN) when modelling nonlinear and high-dimensional data. To this end, the proposed multi-cube unit with sparse attention and ridge regularisation (MCU-SAR) method integrates three complementary components: (i) a multi-cube unit (MCU) architecture that explicitly encodes higher-order feature interactions, (ii) a sparse attention mechanism that suppresses low-informative multiplicative terms, and (iii) a ridge-regularised extreme learning machine (ELM) output layer to improve generalisation. The proposed model is evaluated on 25 publicly available datasets, including 17 classification and 8 regression tasks, and benchmarked against 15 baseline methods comprising gradient-based optimisation techniques, support vector machines (SVM), and various ELM-based approaches. Performance comparisons are conducted using the Friedman test. MCU-SAR demonstrates consistently strong performance, ranking first on the majority of the 25 benchmark datasets and achieving competitive accuracy in classification as well as low error levels in regression tasks, with all results supported by statistically significant p-values. These results demonstrate that the proposed framework provides a scalable, generalisable, and computationally efficient solution for both classification and regression problems, offering robust performance on engineering-oriented real-world datasets.

本研究介绍了一种学习框架，旨在增强单层前馈网络（SLFN）在建模非线性和高维数据时的表示能力和稳定性。为此，提出的具有稀疏注意和脊正则化（MCU- sar）方法的多立方体单元集成了三个互补组件：(i)明确编码高阶特征交互的多立方体单元（MCU）架构，（ii）抑制低信息乘法项的稀疏注意机制，以及（iii）脊正则化极限学习机（ELM）输出层以提高泛化。该模型在25个公开可用的数据集上进行了评估，包括17个分类和8个回归任务，并与15种基线方法进行了基准测试，这些方法包括基于梯度的优化技术、支持向量机（SVM）和各种基于elm的方法。使用弗里德曼测试进行性能比较。MCU-SAR表现出一贯强劲的性能，在25个基准数据集中排名第一，在分类方面具有竞争力的准确性，在回归任务中具有较低的误差水平，所有结果都得到统计显著p值的支持。这些结果表明，所提出的框架为分类和回归问题提供了一个可扩展的、通用的、计算效率高的解决方案，在面向工程的现实世界数据集上提供了强大的性能。

{"title":"Novel multi-cube extreme learning machine with sparse attention and ridge regression technique for robust machine learning","authors":"Merve Zirekgur , Barış Karakaya , Abdulkadir Sengur , U. Rajendra Acharya","doi":"10.1016/j.asoc.2026.114704","DOIUrl":"10.1016/j.asoc.2026.114704","url":null,"abstract":"<div><div>This study introduces a learning framework designed to enhance the representational capacity and stability of single-layer feed-forward networks (SLFN) when modelling nonlinear and high-dimensional data. To this end, the proposed multi-cube unit with sparse attention and ridge regularisation (MCU-SAR) method integrates three complementary components: (i) a multi-cube unit (MCU) architecture that explicitly encodes higher-order feature interactions, (ii) a sparse attention mechanism that suppresses low-informative multiplicative terms, and (iii) a ridge-regularised extreme learning machine (ELM) output layer to improve generalisation. The proposed model is evaluated on 25 publicly available datasets, including 17 classification and 8 regression tasks, and benchmarked against 15 baseline methods comprising gradient-based optimisation techniques, support vector machines (SVM), and various ELM-based approaches. Performance comparisons are conducted using the Friedman test. MCU-SAR demonstrates consistently strong performance, ranking first on the majority of the 25 benchmark datasets and achieving competitive accuracy in classification as well as low error levels in regression tasks, with all results supported by statistically significant p-values. These results demonstrate that the proposed framework provides a scalable, generalisable, and computationally efficient solution for both classification and regression problems, offering robust performance on engineering-oriented real-world datasets.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114704"},"PeriodicalIF":6.6,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080040","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Web data analysis using a hybrid approach of DOM processing and deep learning models 使用DOM处理和深度学习模型的混合方法进行Web数据分析

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-22 DOI: 10.1016/j.asoc.2026.114651

Ba-Vinh Truong , Phu Pham , Loan T.T. Nguyen , Ngoc Thanh Nguyen , Bay Vo

Web data extraction remains a major challenge due to the complexity and variety of Document Object Model (DOM) structures. Previous methods, such as LANTERN and DOM2R-Graph, utilize DOM structures but often struggle to adapt to dynamic or heterogeneous layouts, while general-purpose large language models (LLMs) usually ignore explicit DOM cues and are costly to deploy at scale. This paper introduces DOM-BERT, a hybrid model that combines Bidirectional Encoder Representations from Transformers (BERT) self-attention with compact DOM feature encodings (including tag type, depth, block and XPath/anchor signals), along with a lighter DOM-LSTM baseline based on Long Short-Term Memory (LSTM) for sequential analysis. DOM-BERT jointly captures textual semantics and DOM structure, allowing it to focus on structurally important areas while filtering out noisy content. The model is tested on the SWDE benchmark dataset (124,000 web pages across eight domains). On this benchmark, DOM-BERT attains an average page-level all-fields F1 score of 96.25%, surpassing the DOM-LSTM baseline (92.24%), LANTERN (91.63%) and DOM2R-Graph (93.19%), while matching the accuracy of strong frozen 7–8B-parameter instruction-tuned large language models (96.58%) with roughly two orders of magnitude fewer parameters and much lower training and inference costs.

由于文档对象模型（Document Object Model， DOM）结构的复杂性和多样性，Web数据提取仍然是一个主要挑战。以前的方法，如LANTERN和DOM2R-Graph，利用DOM结构，但往往难以适应动态或异构布局，而通用大型语言模型（llm）通常忽略显式DOM线索，并且大规模部署成本很高。本文介绍了DOM-BERT，这是一个混合模型，它结合了来自变形器（BERT）自关注的双向编码器表示与紧凑的DOM特征编码（包括标签类型、深度、块和XPath/锚信号），以及基于长短期记忆（LSTM）的较轻的DOM-LSTM基线，用于序列分析。DOM- bert联合捕获文本语义和DOM结构，允许它专注于结构上重要的区域，同时过滤掉嘈杂的内容。该模型在SWDE基准数据集（跨八个域的124,000个网页）上进行了测试。在这个基准测试中，DOM-BERT的平均页面级全字段F1得分为96.25%，超过了DOM-LSTM基线（92.24%）、LANTERN（91.63%）和DOM2R-Graph(93.19%)，同时与强冻结的7 - 8b参数指令调优的大型语言模型（96.58%）的准确率相当，参数大约少了两个数量级，训练和推理成本也低得多。

{"title":"Web data analysis using a hybrid approach of DOM processing and deep learning models","authors":"Ba-Vinh Truong , Phu Pham , Loan T.T. Nguyen , Ngoc Thanh Nguyen , Bay Vo","doi":"10.1016/j.asoc.2026.114651","DOIUrl":"10.1016/j.asoc.2026.114651","url":null,"abstract":"<div><div>Web data extraction remains a major challenge due to the complexity and variety of Document Object Model (DOM) structures. Previous methods, such as LANTERN and DOM2R-Graph, utilize DOM structures but often struggle to adapt to dynamic or heterogeneous layouts, while general-purpose large language models (LLMs) usually ignore explicit DOM cues and are costly to deploy at scale. This paper introduces DOM-BERT, a hybrid model that combines Bidirectional Encoder Representations from Transformers (BERT) self-attention with compact DOM feature encodings (including tag type, depth, block and XPath/anchor signals), along with a lighter DOM-LSTM baseline based on Long Short-Term Memory (LSTM) for sequential analysis. DOM-BERT jointly captures textual semantics and DOM structure, allowing it to focus on structurally important areas while filtering out noisy content. The model is tested on the SWDE benchmark dataset (124,000 web pages across eight domains). On this benchmark, DOM-BERT attains an average page-level all-fields F1 score of 96.25%, surpassing the DOM-LSTM baseline (92.24%), LANTERN (91.63%) and DOM2R-Graph (93.19%), while matching the accuracy of strong frozen 7–8B-parameter instruction-tuned large language models (96.58%) with roughly two orders of magnitude fewer parameters and much lower training and inference costs.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114651"},"PeriodicalIF":6.6,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080111","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Spatiotemporal feature fusion modeling method of plantar pressure for pedestrian identity recognition 行人身份识别足底压力时空特征融合建模方法

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-22 DOI: 10.1016/j.asoc.2026.114693

Xinyuan Wei , Qiuyuan Wu , Yanzhu Chang , Nan Zhang , Qiaosheng Pan

Gait contains a unique biometric modality, which is increasingly used for biometric identification, especially in fields such as counter-terrorism and forensics. However, existing identity recognition technology based on plantar pressure remain inadequate for fully capturing the spatial and temporal characteristics of plantar pressure signals, resulting in limited recognition performance and low generalization. To address these limitations, this study proposes a novel multi-scale temporal fusion framework, which is called MCTBA-Net. The proposed framework comprises four complementary modules: Multi-Scale Convolutional Neural Network (MS-CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Temporal Convolutional Network (TCN), and Multi-Head Self-Attention (MHSA). First, the spatial richness of plantar pressure data is captured through the MS-CNN module, which extracts discriminative spatial features across shallow to deep levels using hierarchical multi-scale convolutions. Second, temporal dependencies are modeled using a TCN for short-term dynamics and a Bi-LSTM for long-term sequence learning. Third, the MHSA mechanism is employed to emphasize informative time steps and suppress irrelevant noise. Experimental evaluations on plantar pressure datasets from 100 diverse subjects show that the proposed MCTBA-Net achieves 98.50 % recognition accuracy, significantly outperforming all ablated variants. These results show that the proposed framework is effective in extracting discriminative spatiotemporal features, which has broad applications in gait-based identity recognition.

步态包含一种独特的生物识别模式，越来越多地用于生物识别，特别是在反恐和法医等领域。然而，现有的基于足底压力的身份识别技术还不能充分捕捉足底压力信号的时空特征，识别性能有限，泛化程度较低。为了解决这些限制，本研究提出了一种新的多尺度时间融合框架，称为MCTBA-Net。该框架包括四个互补模块：多尺度卷积神经网络（MS-CNN）、双向长短期记忆（Bi-LSTM）、时间卷积网络（TCN）和多头自注意（MHSA）。首先，通过MS-CNN模块捕获足底压力数据的空间丰富度，该模块使用分层多尺度卷积提取浅层到深层的判别空间特征。其次，使用TCN进行短期动态建模，使用Bi-LSTM进行长期序列学习。第三，利用MHSA机制强调信息时间步长，抑制无关噪声。对来自100个不同受试者的足底压力数据集的实验评估表明，所提出的MCTBA-Net的识别准确率达到98.50 %，显著优于所有消隐变体。实验结果表明，该框架能够有效地提取具有区别性的时空特征，在基于步态的身份识别中具有广泛的应用前景。

{"title":"Spatiotemporal feature fusion modeling method of plantar pressure for pedestrian identity recognition","authors":"Xinyuan Wei , Qiuyuan Wu , Yanzhu Chang , Nan Zhang , Qiaosheng Pan","doi":"10.1016/j.asoc.2026.114693","DOIUrl":"10.1016/j.asoc.2026.114693","url":null,"abstract":"<div><div>Gait contains a unique biometric modality, which is increasingly used for biometric identification, especially in fields such as counter-terrorism and forensics. However, existing identity recognition technology based on plantar pressure remain inadequate for fully capturing the spatial and temporal characteristics of plantar pressure signals, resulting in limited recognition performance and low generalization. To address these limitations, this study proposes a novel multi-scale temporal fusion framework, which is called MCTBA-Net. The proposed framework comprises four complementary modules: Multi-Scale Convolutional Neural Network (MS-CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Temporal Convolutional Network (TCN), and Multi-Head Self-Attention (MHSA). First, the spatial richness of plantar pressure data is captured through the MS-CNN module, which extracts discriminative spatial features across shallow to deep levels using hierarchical multi-scale convolutions. Second, temporal dependencies are modeled using a TCN for short-term dynamics and a Bi-LSTM for long-term sequence learning. Third, the MHSA mechanism is employed to emphasize informative time steps and suppress irrelevant noise. Experimental evaluations on plantar pressure datasets from 100 diverse subjects show that the proposed MCTBA-Net achieves 98.50 % recognition accuracy, significantly outperforming all ablated variants. These results show that the proposed framework is effective in extracting discriminative spatiotemporal features, which has broad applications in gait-based identity recognition.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114693"},"PeriodicalIF":6.6,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039856","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

An ordered weighted averaging-based framework to build and select better composite indicators 一个基于有序加权平均的框架来构建和选择更好的综合指标

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-22 DOI: 10.1016/j.asoc.2026.114710

Matheus Pereira Libório , Petr Iakovlevich Ekel , Marcos Flávio Silveira Vasconcelos D´Angelo , Luis Martínez López , Witold Pedrycz

Social, economic, business, and environmental phenomena are multidimensional in nature. In particular, poverty, economic development, innovation, and sustainability cannot be measured by a single indicator. In this regard, the composite indicators framework provides tools for addressing the multidimensional nature of phenomena. Its implementation offers Polycemaker a one-dimensional synthesis measure, easy to understand, that encompasses the multiple underlying indicators of the multidimensional phenomenon, without loss of analytical scope. Despite its advantages, composite indicators have drawbacks. There are many methods for constructing composite indicators, but some methods remain underexplored. The different ways of constructing composite indicators lead to uncertainty about which method to use. Uncertainty analyses and linkages with external variables have predominantly verified the quality of the composite indicators. This study presents a Software and a Decision Support System to address these gaps. The novel Ordered Weighted Averaging (OWA)-based Software facilitates the construction of composite indicators, leveraging OWA's flexibility to control the sub-indicators' mutual compensation levels and bias composite indicator scores in different directions (positive or negative) and intensities. The Software includes a wide range of tests (robustness, explanatory power, informational power, discriminating power, and the ratio of atypical measurements) to measure the quality of OWA-based composite indicators. Furthermore, the novel Decision Support System enables identifying the composite indicator that achieves higher satisfaction levels in quality tests, preventing decision-makers from using weak solutions. Finally, the Software and the Decision Support System are applied to identify the composite indicator that best represents Global Innovation, Potential Economic Development, and Affordable and Clean Energy.

社会、经济、商业和环境现象在本质上是多维的。特别是，贫困、经济发展、创新和可持续性不能用单一指标来衡量。在这方面，复合指标框架为处理现象的多层面性质提供了工具。它的实施为Polycemaker提供了一个易于理解的一维合成测量，它包含了多维现象的多个潜在指标，而不会失去分析范围。尽管有优点，但综合指标也有缺点。构建复合指标的方法有很多，但有些方法还没有得到充分的探索。构建复合指标的不同方法导致了使用哪种方法的不确定性。不确定性分析和与外部变量的联系主要验证了复合指标的质量。本研究提出了一个软件和决策支持系统来解决这些差距。新颖的基于OWA （Ordered Weighted Averaging）的软件简化了复合指标的构建，利用OWA的灵活性控制子指标的相互补偿水平，并在不同方向（正或负）和强度上偏置综合指标得分。该软件包括广泛的测试（稳健性、解释力、信息能力、判别能力和非典型测量的比率），以衡量基于owa的复合指标的质量。此外，新的决策支持系统能够识别复合指标，在质量测试中达到更高的满意度水平，防止决策者使用弱解决方案。最后，应用软件和决策支持系统来确定最能代表全球创新、潜在经济发展和负担得起的清洁能源的综合指标。

{"title":"An ordered weighted averaging-based framework to build and select better composite indicators","authors":"Matheus Pereira Libório , Petr Iakovlevich Ekel , Marcos Flávio Silveira Vasconcelos D´Angelo , Luis Martínez López , Witold Pedrycz","doi":"10.1016/j.asoc.2026.114710","DOIUrl":"10.1016/j.asoc.2026.114710","url":null,"abstract":"<div><div>Social, economic, business, and environmental phenomena are multidimensional in nature. In particular, poverty, economic development, innovation, and sustainability cannot be measured by a single indicator. In this regard, the composite indicators framework provides tools for addressing the multidimensional nature of phenomena. Its implementation offers Polycemaker a one-dimensional synthesis measure, easy to understand, that encompasses the multiple underlying indicators of the multidimensional phenomenon, without loss of analytical scope. Despite its advantages, composite indicators have drawbacks. There are many methods for constructing composite indicators, but some methods remain underexplored. The different ways of constructing composite indicators lead to uncertainty about which method to use. Uncertainty analyses and linkages with external variables have predominantly verified the quality of the composite indicators. This study presents a Software and a Decision Support System to address these gaps. The novel Ordered Weighted Averaging (OWA)-based Software facilitates the construction of composite indicators, leveraging OWA's flexibility to control the sub-indicators' mutual compensation levels and bias composite indicator scores in different directions (positive or negative) and intensities. The Software includes a wide range of tests (robustness, explanatory power, informational power, discriminating power, and the ratio of atypical measurements) to measure the quality of OWA-based composite indicators. Furthermore, the novel Decision Support System enables identifying the composite indicator that achieves higher satisfaction levels in quality tests, preventing decision-makers from using weak solutions. Finally, the Software and the Decision Support System are applied to identify the composite indicator that best represents Global Innovation, Potential Economic Development, and Affordable and Clean Energy.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114710"},"PeriodicalIF":6.6,"publicationDate":"2026-01-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080302","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A hybrid agent based on reinforcement learning and fuzzy computing using q-rung orthopair fuzzy information for electricity purchasing negotiation in smart grids 基于q阶正交模糊信息的强化学习和模糊计算混合智能体用于智能电网购电协商

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-21 DOI: 10.1016/j.asoc.2026.114683

Dimitrios K. Panagiotou, Anastasios Dounis

Smart Grid electricity procurement requires efficient multi-issue negotiation under uncertainty, where participants lack complete information about opponents’ preferences and must balance individual utility with social welfare. This paper proposes a hybrid negotiation framework that integrates q-rung orthopair fuzzy sets (q-ROFS) for preference modeling, Q-learning for online estimation of opponent weights, and fuzzy inference systems for adaptive concessions and acceptance decisions. Buyer offers are ranked using a q-ROFS aggregation mechanism that captures preference vagueness, while reinforcement learning dynamically estimates opponent issue weights based on observed negotiation outcomes and Kaldor-Hicks welfare changes. The proposed model is evaluated in a simulated electricity purchasing scenario involving multiple sellers and three negotiation issues across different q values. Results demonstrate that the proposed approach consistently improves joint utility and fairness compared to a Stackelberg-inspired benchmark, achieving an average increase in joint utility of 3.8 % while producing more balanced and robust agreements that approach the Pareto frontier. These findings confirm the effectiveness of combining fuzzy preference modeling, learning-based adaptation, and welfare-oriented rewards in automated energy negotiations.

智能电网电力采购需要在不确定条件下进行高效的多议题谈判，参与者缺乏对手偏好的完全信息，必须平衡个人效用与社会福利。本文提出了一种混合协商框架，该框架集成了用于偏好建模的q-rung正形模糊集（q-ROFS）、用于在线估计对手权重的q-学习以及用于自适应让步和接受决策的模糊推理系统。买方报价使用捕获偏好模糊的q-ROFS聚合机制进行排名，而强化学习基于观察到的谈判结果和卡尔多-希克斯福利变化动态估计对手问题权重。在一个涉及多个卖家和三个不同q值的谈判问题的模拟购电场景中，对所提出的模型进行了评估。结果表明，与stackelberg启发的基准相比，所提出的方法持续提高了联合效用和公平性，实现了3.8 %的联合效用平均增长，同时产生了更平衡和更稳健的协议，接近帕累托边界。这些研究结果证实了模糊偏好建模、基于学习的适应和基于福利的奖励相结合在自动化能源谈判中的有效性。

{"title":"A hybrid agent based on reinforcement learning and fuzzy computing using q-rung orthopair fuzzy information for electricity purchasing negotiation in smart grids","authors":"Dimitrios K. Panagiotou, Anastasios Dounis","doi":"10.1016/j.asoc.2026.114683","DOIUrl":"10.1016/j.asoc.2026.114683","url":null,"abstract":"<div><div>Smart Grid electricity procurement requires efficient multi-issue negotiation under uncertainty, where participants lack complete information about opponents’ preferences and must balance individual utility with social welfare. This paper proposes a hybrid negotiation framework that integrates q-rung orthopair fuzzy sets (q-ROFS) for preference modeling, Q-learning for online estimation of opponent weights, and fuzzy inference systems for adaptive concessions and acceptance decisions. Buyer offers are ranked using a q-ROFS aggregation mechanism that captures preference vagueness, while reinforcement learning dynamically estimates opponent issue weights based on observed negotiation outcomes and Kaldor-Hicks welfare changes. The proposed model is evaluated in a simulated electricity purchasing scenario involving multiple sellers and three negotiation issues across different q values. Results demonstrate that the proposed approach consistently improves joint utility and fairness compared to a Stackelberg-inspired benchmark, achieving an average increase in joint utility of 3.8 % while producing more balanced and robust agreements that approach the Pareto frontier. These findings confirm the effectiveness of combining fuzzy preference modeling, learning-based adaptation, and welfare-oriented rewards in automated energy negotiations.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114683"},"PeriodicalIF":6.6,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146039864","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

An adaptive exploration-exploitation framework for success history based differential evolution 基于差异演化成功史的自适应勘探开发框架

IF 6.6 1区计算机科学 Q1 COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE

Applied Soft Computing

Pub Date : 2026-01-21 DOI: 10.1016/j.asoc.2026.114665

Yiwen Zhuo , Qiangda Yang

Differential evolution (DE) algorithms are widely applied to various optimization problems due to their simplicity and excellent performance. Among these, Success-History based Adaptive DE (SHADE) and its variants have gained significant attention and shown high competitiveness owing to the robustness and superior performance enabled by their parameter adaptation strategies. However, when confronted with multimodal problems, these conventional parameter adaptation strategies often fail to generate appropriate parameters to guide the population's exploration. This readily leads to an imbalance between exploration and exploitation, ultimately degrading performance. To address this limitation, we propose a parameter configuration framework that extends conventional adaptation strategies. Designed for broad applicability, this framework is readily extensible to all major SHADE-based algorithms. It enhances parameter settings in these algorithms via a semi-parameter adaptation strategy that guides population exploration toward promising regions, thereby overcoming the limitations of conventional parameter adaptation in multimodal optimization. Furthermore, we introduce a dual-criterion identification mechanism based on offspring success rate and population diversity metrics. Integrated within the aforementioned strategy, this mechanism facilitates adaptive exploration-exploitation balancing. Full testing across the IEEE Congress on Evolutionary Computation (CEC) 2014, 2015, and 2017 benchmark suites demonstrates that applying the framework comprising the aforementioned strategy and mechanism to both competition-winning and recently published SHADE-based algorithms significantly boosts their performance over the original versions. Finally, we discuss parameter sensitivity and algorithmic complexity, and validate the effectiveness through visualization analysis.

差分进化算法以其简单和优异的性能被广泛应用于各种优化问题。其中，基于成功历史的自适应DE （SHADE）及其变体由于其参数自适应策略的鲁棒性和卓越性能而获得了极大的关注，并表现出很高的竞争力。然而，当面对多模态问题时，这些传统的参数自适应策略往往不能产生合适的参数来指导群体的探索。这很容易导致勘探和开采之间的不平衡，最终降低性能。为了解决这一限制，我们提出了一个扩展传统适应策略的参数配置框架。为广泛的适用性而设计，这个框架很容易扩展到所有主要的基于shade的算法。该算法通过半参数自适应策略增强了算法的参数设置，引导种群向有希望的区域探索，从而克服了传统参数自适应在多模态优化中的局限性。此外，我们还引入了一种基于后代成功率和种群多样性指标的双标准识别机制。在前面提到的策略中，这种机制促进了适应性的探索-开发平衡。在IEEE进化计算大会（CEC） 2014年、2015年和2017年的基准测试套件中进行的全面测试表明，将包含上述策略和机制的框架应用于竞争获胜和最近发布的基于shade的算法，显著提高了其性能。最后讨论了参数敏感性和算法复杂度，并通过可视化分析验证了算法的有效性。

{"title":"An adaptive exploration-exploitation framework for success history based differential evolution","authors":"Yiwen Zhuo , Qiangda Yang","doi":"10.1016/j.asoc.2026.114665","DOIUrl":"10.1016/j.asoc.2026.114665","url":null,"abstract":"<div><div>Differential evolution (DE) algorithms are widely applied to various optimization problems due to their simplicity and excellent performance. Among these, Success-History based Adaptive DE (SHADE) and its variants have gained significant attention and shown high competitiveness owing to the robustness and superior performance enabled by their parameter adaptation strategies. However, when confronted with multimodal problems, these conventional parameter adaptation strategies often fail to generate appropriate parameters to guide the population's exploration. This readily leads to an imbalance between exploration and exploitation, ultimately degrading performance. To address this limitation, we propose a parameter configuration framework that extends conventional adaptation strategies. Designed for broad applicability, this framework is readily extensible to all major SHADE-based algorithms. It enhances parameter settings in these algorithms via a semi-parameter adaptation strategy that guides population exploration toward promising regions, thereby overcoming the limitations of conventional parameter adaptation in multimodal optimization. Furthermore, we introduce a dual-criterion identification mechanism based on offspring success rate and population diversity metrics. Integrated within the aforementioned strategy, this mechanism facilitates adaptive exploration-exploitation balancing. Full testing across the IEEE Congress on Evolutionary Computation (CEC) 2014, 2015, and 2017 benchmark suites demonstrates that applying the framework comprising the aforementioned strategy and mechanism to both competition-winning and recently published SHADE-based algorithms significantly boosts their performance over the original versions. Finally, we discuss parameter sensitivity and algorithmic complexity, and validate the effectiveness through visualization analysis.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114665"},"PeriodicalIF":6.6,"publicationDate":"2026-01-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080100","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0