Applied Soft Computing最新文献_第3页

Incorporating mixed-effects into Adaptive time relation multi-task learning with longitudinal data for Alzheimer’s disease progression prediction 基于纵向数据的自适应时间关系多任务学习混合效应在阿尔茨海默病进展预测中的应用

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

Applied Soft Computing

Pub Date : 2026-01-24 DOI: 10.1016/j.asoc.2026.114716

Linting Miao , Hongmei Chen , Biao Xiang , Zhong Yuan , Chuan Luo , Tianrui Li

Alzheimer’s disease (AD) is a slowly progressing neurodegenerative disorder and poses significant challenges for early diagnosis and longitudinal prognosis in the medical sector. Accurate prediction of cognitive decline is crucial for timely clinical intervention, disease monitoring, and treatment planning. Multi-task learning (MTL) has been extensively applied in AD prediction tasks, as it effectively captures shared patterns across multiple objectives and improves generalization. However, most existing MTL-based approaches focus on cross-sectional settings and lack the ability to explicitly model disease progression over time. To address this limitation, we propose a longitudinal multi-task learning framework that jointly models disease progression and adaptive temporal relationships using multi-timepoint neuroimaging data. The proposed method incorporates two task-specific projection matrices within a mixed-effects modeling framework to disentangle baseline-invariant representations from change-sensitive features, thereby capturing distinct patterns attributable to AD pathology and normal aging. Temporal relationships among tasks are learned directly from data via a task relationship matrix, while temporal asymmetry is enforced through directional regularization. Structured regularization is further introduced to enhance the sparsity and robustness of the learned model. The proposed framework is evaluated on real-world datasets from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) using standard regression metrics, including root mean squared error (rMSE). Compared with the best-performing baselines, our model achieves an average rMSE reduction of approximately 6%–10% across three widely used cognitive measures at multiple time points, with improvements validated by statistical significance testing, indicating more accurate and reliable prediction of cognitive decline. Beyond predictive accuracy, the model also provides enhanced interpretability through brain-region-level visualization, which facilitates a clearer understanding of disease-related progression patterns and age-related effects, and supports clinical analysis and decision-making.

阿尔茨海默病（AD）是一种进展缓慢的神经退行性疾病，在医学领域对早期诊断和纵向预后提出了重大挑战。准确预测认知能力下降对于及时的临床干预、疾病监测和治疗计划至关重要。多任务学习（Multi-task learning， MTL）在AD预测任务中得到了广泛的应用，因为它可以有效地捕获跨多个目标的共享模式并提高泛化能力。然而，大多数现有的基于mtl的方法侧重于横断面设置，缺乏明确模拟疾病随时间进展的能力。为了解决这一限制，我们提出了一个纵向多任务学习框架，该框架使用多时间点神经成像数据联合建模疾病进展和自适应时间关系。该方法在混合效应建模框架中结合了两个特定任务的投影矩阵，以从变化敏感特征中分离出基线不变表示，从而捕获可归因于AD病理和正常衰老的不同模式。任务之间的时间关系通过任务关系矩阵直接从数据中学习，而时间不对称通过定向正则化来实现。进一步引入结构化正则化来增强学习模型的稀疏性和鲁棒性。使用标准回归指标（包括均方根误差（rMSE））对来自阿尔茨海默病神经影像学倡议（ADNI）的真实数据集进行评估。与表现最好的基线相比，我们的模型在多个时间点上实现了三种广泛使用的认知测量的平均rMSE降低约6%-10%，并通过统计显著性检验验证了改进，表明对认知衰退的预测更准确、更可靠。除了预测准确性之外，该模型还通过大脑区域级别的可视化提供了增强的可解释性，这有助于更清楚地了解疾病相关的进展模式和年龄相关的影响，并支持临床分析和决策。

{"title":"Incorporating mixed-effects into Adaptive time relation multi-task learning with longitudinal data for Alzheimer’s disease progression prediction","authors":"Linting Miao , Hongmei Chen , Biao Xiang , Zhong Yuan , Chuan Luo , Tianrui Li","doi":"10.1016/j.asoc.2026.114716","DOIUrl":"10.1016/j.asoc.2026.114716","url":null,"abstract":"<div><div>Alzheimer’s disease (AD) is a slowly progressing neurodegenerative disorder and poses significant challenges for early diagnosis and longitudinal prognosis in the medical sector. Accurate prediction of cognitive decline is crucial for timely clinical intervention, disease monitoring, and treatment planning. Multi-task learning (MTL) has been extensively applied in AD prediction tasks, as it effectively captures shared patterns across multiple objectives and improves generalization. However, most existing MTL-based approaches focus on cross-sectional settings and lack the ability to explicitly model disease progression over time. To address this limitation, we propose a longitudinal multi-task learning framework that jointly models disease progression and adaptive temporal relationships using multi-timepoint neuroimaging data. The proposed method incorporates two task-specific projection matrices within a mixed-effects modeling framework to disentangle baseline-invariant representations from change-sensitive features, thereby capturing distinct patterns attributable to AD pathology and normal aging. Temporal relationships among tasks are learned directly from data via a task relationship matrix, while temporal asymmetry is enforced through directional regularization. Structured regularization is further introduced to enhance the sparsity and robustness of the learned model. The proposed framework is evaluated on real-world datasets from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) using standard regression metrics, including root mean squared error (rMSE). Compared with the best-performing baselines, our model achieves an average rMSE reduction of approximately 6%–10% across three widely used cognitive measures at multiple time points, with improvements validated by statistical significance testing, indicating more accurate and reliable prediction of cognitive decline. Beyond predictive accuracy, the model also provides enhanced interpretability through brain-region-level visualization, which facilitates a clearer understanding of disease-related progression patterns and age-related effects, and supports clinical analysis and decision-making.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"192 ","pages":"Article 114716"},"PeriodicalIF":6.6,"publicationDate":"2026-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146096143","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

SE-RSRNet: Rain streak removal by feature selection and feature extraction SE-RSRNet：基于特征选择和特征提取的雨纹去除

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

Applied Soft Computing

Pub Date : 2026-01-24 DOI: 10.1016/j.asoc.2026.114702

Chuang Chen , Xiaopei Zhang , Yumeng Niu , Bo Li , Luoyu Zhou

Mine tunnel images are often affected by underground water leakage, resulting in rain-like streaks that degrade visual quality and seriously interfere with automated monitoring and perception tasks. However, due to the extremely low illumination, heavy dust, weak contrast, and simple structural textures in underground environments, existing deraining methods perform poorly and exhibit clear limitations in such challenging conditions. To address these issues, this paper proposes a novel rain streak removal network (SE-RSRNet), designed to effectively eliminate rain streaks while preserving background textures and fine visual details. The model introduces a GLU (Gated Linear Unit) module to adaptively select critical features through a gating mechanism, followed by a U-FFA (U-Net Feature Fusion Attention) module to improve multi-scale feature fusion and detail restoration. Additionally, a WCAB (Weighted Channel Attention Block) is incorporated to further enhance feature representation through refined channel attention. Extensive ablation experiments verify the effectiveness of each module within the architecture. Compared with the baseline model, the proposed SE-RSRNet improves PSNR by 0.41 dB. On the Rain100L dataset, it achieves a PSNR gain of 0.61 dB over other state-of-the-art methods, demonstrating superior performance. Furthermore, applying the derained images to a YOLOv5s-based detection system in real mine monitoring scenarios significantly improves target detection accuracy, confirming the practical value of the proposed approach.

矿井隧道图像经常受到地下漏水的影响，产生类似雨的条纹，降低了视觉质量，严重干扰了自动监测和感知任务。然而，由于地下环境光照极低、粉尘大、对比度弱、结构纹理简单，现有的脱模方法在这种具有挑战性的条件下表现不佳，局限性明显。为了解决这些问题，本文提出了一种新的雨纹去除网络（SE-RSRNet），旨在有效地去除雨纹，同时保留背景纹理和良好的视觉细节。该模型引入了GLU（门控线性单元）模块，通过门控机制自适应选择关键特征，然后引入U-FFA （U-Net特征融合注意）模块，以提高多尺度特征融合和细节恢复。此外，还引入了加权信道注意块（WCAB），通过改进信道注意进一步增强特征表示。大量的烧蚀实验验证了体系结构中每个模块的有效性。与基线模型相比，SE-RSRNet的PSNR提高了0.41 dB。在Rain100L数据集上，与其他最先进的方法相比，它实现了0.61 dB的PSNR增益，显示出优越的性能。此外，将提取的图像应用于基于yolov5的探测系统，在真实的矿山监测场景中，显著提高了目标探测精度，验证了所提方法的实用价值。

{"title":"SE-RSRNet: Rain streak removal by feature selection and feature extraction","authors":"Chuang Chen , Xiaopei Zhang , Yumeng Niu , Bo Li , Luoyu Zhou","doi":"10.1016/j.asoc.2026.114702","DOIUrl":"10.1016/j.asoc.2026.114702","url":null,"abstract":"<div><div>Mine tunnel images are often affected by underground water leakage, resulting in rain-like streaks that degrade visual quality and seriously interfere with automated monitoring and perception tasks. However, due to the extremely low illumination, heavy dust, weak contrast, and simple structural textures in underground environments, existing deraining methods perform poorly and exhibit clear limitations in such challenging conditions. To address these issues, this paper proposes a novel rain streak removal network (SE-RSRNet), designed to effectively eliminate rain streaks while preserving background textures and fine visual details. The model introduces a GLU (Gated Linear Unit) module to adaptively select critical features through a gating mechanism, followed by a U-FFA (U-Net Feature Fusion Attention) module to improve multi-scale feature fusion and detail restoration. Additionally, a WCAB (Weighted Channel Attention Block) is incorporated to further enhance feature representation through refined channel attention. Extensive ablation experiments verify the effectiveness of each module within the architecture. Compared with the baseline model, the proposed SE-RSRNet improves PSNR by 0.41 dB. On the Rain100L dataset, it achieves a PSNR gain of 0.61 dB over other state-of-the-art methods, demonstrating superior performance. Furthermore, applying the derained images to a YOLOv5s-based detection system in real mine monitoring scenarios significantly improves target detection accuracy, confirming the practical value of the proposed approach.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114702"},"PeriodicalIF":6.6,"publicationDate":"2026-01-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080196","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

MFU-Depth: Modeling and fusing uncertainty for self-supervised multi-frame monocular depth estimation MFU-Depth：自监督多帧单目深度估计的建模和融合不确定性

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.2026.114705

Fuji Fu , Jinfu Yang , Jiaqi Ma

Self-supervised monocular depth estimation, valued for its simplified configuration and excellent geometric perception, is gaining attention in autonomous driving and robotics. Mainstream self-supervised multi-frame methods, due to the lack of a unified multi-cue uncertainty representation mechanism, often struggle to effectively reflect depth prediction errors. To address this issue, we propose a self-supervised multi-frame monocular depth estimation method by Modeling and Fusing Uncertainty, termed MFU-Depth. (1) At the probability distribution level, we design an Uncertainty Estimation-based Single-frame Probability Prediction (UE-SPP) module and a Depth-Aware Multi-frame Probability Transformation (DA-MPT) module, achieving unified uncertainty modeling of single-frame and multi-frame cues in the form of probability distributions. Building on these, we establish a Relation Modeling-guided Probability Fusion (RM-PF) mechanism, which adaptively adjusts fusion weights through pixel-wise difference relation modeling, enabling the complementary integration of these two distributions. (2) At the probabilistic self-supervision level, an Information Entropy Uncertainty Loss (IEUL) is devised to further model the uncertainty of both single-frame and fused depth probability distributions, thereby mitigating supervision distortion caused by high-uncertainty pixels. Experimental results show that MFU-Depth achieves exceptional performance on multiple public datasets. Compared to baseline methods, MFU-Depth reduces the depth error by 15.3% and the uncertainty metric by 29.0%.

自监督单目深度估计以其简化的结构和优异的几何感知能力在自动驾驶和机器人领域受到越来越多的关注。主流的自监督多帧方法由于缺乏统一的多线索不确定性表示机制，往往难以有效反映深度预测误差。为了解决这个问题，我们提出了一种基于建模和融合不确定性的自监督多帧单目深度估计方法，称为MFU-Depth。(1)在概率分布层面，设计了基于不确定性估计的单帧概率预测（UE-SPP）模块和深度感知的多帧概率变换（DA-MPT）模块，实现了单帧和多帧线索以概率分布形式的统一不确定性建模。在此基础上，我们建立了一种关系建模引导的概率融合（RM-PF）机制，该机制通过逐像素的差异关系建模自适应调整融合权重，实现了这两种分布的互补集成。(2)在概率自监督层面，设计了一种信息熵不确定性损失（Information Entropy Uncertainty Loss， IEUL），进一步对单帧和融合深度概率分布的不确定性进行建模，从而减轻高不确定性像素造成的监督失真。实验结果表明，MFU-Depth在多个公共数据集上取得了优异的性能。与基线方法相比，MFU-Depth将深度误差降低了15.3%，不确定度度量降低了29.0%。

{"title":"MFU-Depth: Modeling and fusing uncertainty for self-supervised multi-frame monocular depth estimation","authors":"Fuji Fu , Jinfu Yang , Jiaqi Ma","doi":"10.1016/j.asoc.2026.114705","DOIUrl":"10.1016/j.asoc.2026.114705","url":null,"abstract":"<div><div>Self-supervised monocular depth estimation, valued for its simplified configuration and excellent geometric perception, is gaining attention in autonomous driving and robotics. Mainstream self-supervised multi-frame methods, due to the lack of a unified multi-cue uncertainty representation mechanism, often struggle to effectively reflect depth prediction errors. To address this issue, we propose a self-supervised multi-frame monocular depth estimation method by Modeling and Fusing Uncertainty, termed MFU-Depth. (1) At the probability distribution level, we design an Uncertainty Estimation-based Single-frame Probability Prediction (UE-SPP) module and a Depth-Aware Multi-frame Probability Transformation (DA-MPT) module, achieving unified uncertainty modeling of single-frame and multi-frame cues in the form of probability distributions. Building on these, we establish a Relation Modeling-guided Probability Fusion (RM-PF) mechanism, which adaptively adjusts fusion weights through pixel-wise difference relation modeling, enabling the complementary integration of these two distributions. (2) At the probabilistic self-supervision level, an Information Entropy Uncertainty Loss (IEUL) is devised to further model the uncertainty of both single-frame and fused depth probability distributions, thereby mitigating supervision distortion caused by high-uncertainty pixels. Experimental results show that MFU-Depth achieves exceptional performance on multiple public datasets. Compared to baseline methods, MFU-Depth reduces the depth error by 15.3% and the uncertainty metric by 29.0%.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114705"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080105","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

MACSENet: A novel lightweight CNN with multi-scale atrous convolutions and attention mechanism for accurate lung cancer detection 基于多尺度亚特卷积和关注机制的新型轻量级CNN精确肺癌检测

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.2026.114698

Fazal Hadi , Omair Bilal , Sohaib Asif

Lung cancer represents a critical global health challenge, characterized by high mortality rates and complex diagnostic processes. Identifying tumors in CT images manually is a time-intensive and complex task. While deep learning models can aid in the process, their complexity often leads to overfitting, and their single-branch, linear structures struggle to capture multi-scale features, ultimately reducing detection accuracy. This paper presents MACSENet, an innovative lightweight CNN developed to improve early detection and diagnostic accuracy using advanced deep learning techniques. The proposed model combines three key components: a standard CNN for basic feature extraction, a multiscale atrous convolutional module for capturing spatial features at different resolutions, and a squeeze-and-excitation (SE) block to refine channel-wise features through adaptive attention mechanisms. We designed the Multiscale Atrous Feature Extraction Block (MAFEB), which incorporates a multiscale atrous convolution module with varying dilation rates. This approach expands the receptive field without adding extra parameters, capturing intricate tumor details and providing a comprehensive contextual understanding to address the complex morphological variations in lung cancer. The SE block is applied after extracting features from the multiscale atrous convolution module, dynamically refining the feature maps by amplifying key diagnostic information and suppressing irrelevant details. To assess the broader applicability and generalizability of MACSENet, we evaluated it on the publicly available IQ-OTH/NCCD lung cancer dataset, which consists of 1097 CT images from 110 patients, encompassing benign, malignant, and normal cases. Additional experiments were carried out to evaluate the robustness of MACSENet using two other datasets: the Chest CT-Scan dataset, which consists of 1000 images across four classes, and the Lung histopathological images dataset, containing 15,000 images across three classes. The experimental results demonstrate that MACSENet outperforms both traditional deep CNNs and state-of-the-art methods in terms of accuracy and efficiency. Our approach achieved an outstanding accuracy of 99.55 % on the IQ-OTH/NCCD dataset, and 92 % and 99.97 % on the additional datasets while maintaining the smallest parameter size of 0.27 million among all competitors. Additionally, the model exhibited the smallest model size and FLOPS compared to other deep CNNs. Our proposed model is also compared with well-established baseline models including ResNet, DenseNet, and MobileNet series to demonstrate its superior performance in lung cancer detection. The proposed approach provides critical insights to support clinical decision-making for advancing patient care and treatment precision in lung cancer diagnostics.

肺癌是一项重大的全球健康挑战，其特点是死亡率高，诊断过程复杂。人工识别CT图像中的肿瘤是一项耗时且复杂的任务。虽然深度学习模型可以在这一过程中提供帮助，但它们的复杂性往往会导致过拟合，而且它们的单分支线性结构难以捕捉多尺度特征，最终降低了检测精度。本文介绍了MACSENet，一种创新的轻量级CNN，用于使用先进的深度学习技术提高早期检测和诊断的准确性。该模型结合了三个关键组件：用于基本特征提取的标准CNN，用于捕获不同分辨率空间特征的多尺度亚历克斯卷积模块，以及通过自适应注意机制来细化通道特征的挤压和激励（SE）块。我们设计了多尺度亚光束特征提取块（mateb），它包含了一个具有不同膨胀率的多尺度亚光束卷积模块。这种方法在不增加额外参数的情况下扩展了感受野，捕获了复杂的肿瘤细节，并提供了全面的背景理解，以解决肺癌中复杂的形态变化。从多尺度亚属性卷积模块中提取特征，通过放大关键诊断信息和抑制无关细节来动态细化特征映射，然后应用SE块。为了评估MACSENet更广泛的适用性和泛化性，我们在公开的IQ-OTH/NCCD肺癌数据集上对其进行了评估，该数据集由110名患者的1097张CT图像组成，包括良性、恶性和正常病例。使用另外两个数据集进行了额外的实验来评估MACSENet的鲁棒性：胸部ct扫描数据集，包含四个类别的1000张图像，以及肺组织病理学图像数据集，包含三个类别的15,000张图像。实验结果表明，MACSENet在精度和效率方面优于传统的深度cnn和最先进的方法。我们的方法在IQ-OTH/NCCD数据集上取得了99.55 %的出色准确率，在其他数据集上达到了92 %和99.97 %，同时在所有竞争对手中保持了最小的参数大小0.27万。此外，与其他深度cnn相比，该模型具有最小的模型尺寸和FLOPS。我们提出的模型还与已建立的基线模型（包括ResNet， DenseNet和MobileNet系列）进行了比较，以证明其在肺癌检测方面的优越性能。所提出的方法为支持临床决策提供了关键的见解，以提高肺癌诊断的患者护理和治疗精度。

{"title":"MACSENet: A novel lightweight CNN with multi-scale atrous convolutions and attention mechanism for accurate lung cancer detection","authors":"Fazal Hadi , Omair Bilal , Sohaib Asif","doi":"10.1016/j.asoc.2026.114698","DOIUrl":"10.1016/j.asoc.2026.114698","url":null,"abstract":"<div><div>Lung cancer represents a critical global health challenge, characterized by high mortality rates and complex diagnostic processes. Identifying tumors in CT images manually is a time-intensive and complex task. While deep learning models can aid in the process, their complexity often leads to overfitting, and their single-branch, linear structures struggle to capture multi-scale features, ultimately reducing detection accuracy. This paper presents MACSENet, an innovative lightweight CNN developed to improve early detection and diagnostic accuracy using advanced deep learning techniques. The proposed model combines three key components: a standard CNN for basic feature extraction, a multiscale atrous convolutional module for capturing spatial features at different resolutions, and a squeeze-and-excitation (SE) block to refine channel-wise features through adaptive attention mechanisms. We designed the Multiscale Atrous Feature Extraction Block (MAFEB), which incorporates a multiscale atrous convolution module with varying dilation rates. This approach expands the receptive field without adding extra parameters, capturing intricate tumor details and providing a comprehensive contextual understanding to address the complex morphological variations in lung cancer. The SE block is applied after extracting features from the multiscale atrous convolution module, dynamically refining the feature maps by amplifying key diagnostic information and suppressing irrelevant details. To assess the broader applicability and generalizability of MACSENet, we evaluated it on the publicly available IQ-OTH/NCCD lung cancer dataset, which consists of 1097 CT images from 110 patients, encompassing benign, malignant, and normal cases. Additional experiments were carried out to evaluate the robustness of MACSENet using two other datasets: the Chest CT-Scan dataset, which consists of 1000 images across four classes, and the Lung histopathological images dataset, containing 15,000 images across three classes. The experimental results demonstrate that MACSENet outperforms both traditional deep CNNs and state-of-the-art methods in terms of accuracy and efficiency. Our approach achieved an outstanding accuracy of 99.55 % on the IQ-OTH/NCCD dataset, and 92 % and 99.97 % on the additional datasets while maintaining the smallest parameter size of 0.27 million among all competitors. Additionally, the model exhibited the smallest model size and FLOPS compared to other deep CNNs. Our proposed model is also compared with well-established baseline models including ResNet, DenseNet, and MobileNet series to demonstrate its superior performance in lung cancer detection. The proposed approach provides critical insights to support clinical decision-making for advancing patient care and treatment precision in lung cancer diagnostics.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114698"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080113","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

Unsupervised sparse temporal knowledge graph entity alignment via joint temporal relational representation learning 基于联合时态关系表示学习的无监督稀疏时态知识图实体对齐

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.2026.114684

Xuan Du, Luyi Bai, Guangchen Feng, Jingfang Li

Entity Alignment (EA) aims to find entities from different knowledge graphs (KGs) that point to the same object in the real world. In recent years, Temporal Knowledge Graphs (TKGs) have extended static KGs by introducing timestamps, providing a new perspective for EA. However, the current mainstream TKG EA models are supervised models that rely on pre-aligned seeds and implicitly encode temporal information into the entity embedding space for identifying equivalent entities. In addition, the current mainstream TKGs entity alignment methods do not make full use of temporal information and do not consider the sparsity of information. To solve the above challenges, we propose an Unsupervised Sparse Temporal Knowledge Graph Entity Alignment model via joint temporal relational representation learning, namely USTEA. In this framework, we propose multi-layer convolutional propagation to supplement entities lacking temporal information, improving the quality of unsupervised alignment seed pairs on sparse temporal knowledge graphs (STKGs). Moreover, we introduce temporal relational representation learning that effectively captures the sparse temporal and relational information on sparse temporal knowledge graphs (STKGs). Experimental results on four STKGs demonstrate that the USTEA model outperforms both supervised and unsupervised state-of-the-art TKG EA models.

实体对齐（EA）旨在从不同的知识图（KGs）中找到指向现实世界中相同对象的实体。近年来，时间知识图（TKGs）通过引入时间戳对静态知识图进行了扩展，为EA提供了一个新的视角。然而，目前主流的TKG知识图模型是依赖于预排列种子的监督模型，并将时间信息隐式编码到实体嵌入空间中以识别等效实体。此外，目前主流的TKGs实体对齐方法没有充分利用时间信息，也没有考虑信息的稀疏性。为了解决上述问题，我们提出了一种基于联合时态关系表示学习的无监督稀疏时态知识图实体对齐模型，即USTEA。在该框架中，我们提出了多层卷积传播来补充缺乏时间信息的实体，提高了稀疏时间知识图（STKGs）上无监督对齐种子对的质量。此外，我们引入了时间关系表示学习，有效地捕获稀疏时间知识图（STKGs）上的稀疏时间和关系信息。在四个stkg上的实验结果表明，USTEA模型优于有监督和无监督的最先进的TKG EA模型。

{"title":"Unsupervised sparse temporal knowledge graph entity alignment via joint temporal relational representation learning","authors":"Xuan Du, Luyi Bai, Guangchen Feng, Jingfang Li","doi":"10.1016/j.asoc.2026.114684","DOIUrl":"10.1016/j.asoc.2026.114684","url":null,"abstract":"<div><div>Entity Alignment (EA) aims to find entities from different knowledge graphs (KGs) that point to the same object in the real world. In recent years, Temporal Knowledge Graphs (TKGs) have extended static KGs by introducing timestamps, providing a new perspective for EA. However, the current mainstream TKG EA models are supervised models that rely on pre-aligned seeds and implicitly encode temporal information into the entity embedding space for identifying equivalent entities. In addition, the current mainstream TKGs entity alignment methods do not make full use of temporal information and do not consider the sparsity of information. To solve the above challenges, we propose an Unsupervised Sparse Temporal Knowledge Graph Entity Alignment model via joint temporal relational representation learning, namely USTEA. In this framework, we propose multi-layer convolutional propagation to supplement entities lacking temporal information, improving the quality of unsupervised alignment seed pairs on sparse temporal knowledge graphs (STKGs). Moreover, we introduce temporal relational representation learning that effectively captures the sparse temporal and relational information on sparse temporal knowledge graphs (STKGs). Experimental results on four STKGs demonstrate that the USTEA model outperforms both supervised and unsupervised state-of-the-art TKG EA models.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114684"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080195","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

GLUE-GAN: Global-local underwater image enhancement generative adversarial network GLUE-GAN：全局-局部水下图像增强生成对抗网络

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.2026.114686

Jun Wang , Daoyi Wei , Wentao Shi , Chenyang Liu , Wenlong Jiang , Juan Wang

Underwater images suffer from color cast, blur, and low contrast due to wavelength-dependent absorption, scattering, and suspended particulates. Many enhancement methods restore global appearance while suppressing local features; others operate in the frequency domain or rely on auxiliary regularization but lack an explicit mechanism to fuse global and local evidence. We introduce GLUE-GAN, a generative adversarial network that globally–locally unifies enhancement through explicit cross-scale, cross-domain fusion. The model comprises: (1) an adaptive feature enhanced encoder that marries spatial context modules with grouped channel-wise self-attention to collaboratively model spatial–channel dependencies across diverse scenes; (2) a multichannel feature aggregation enhancement module that performs multi-scale extraction and alignment to uniformly recover global tone and local textures; and (3) a global–local information enhancement module that uses wavelet decomposition to separate low- and high-frequency bands, processing that mitigates local bias during global correction. Evaluations on EUVP, UIEB, and UFO-120 demonstrate consistent gains in color fidelity, contrast, and sharpness, with improved preservation of edges and fine details. By unifying spatial–channel reasoning with frequency-aware processing in a single adversarial framework, GLUE-GAN balances global color correction and local detail preservation for underwater image enhancement.

由于波长依赖的吸收、散射和悬浮微粒，水下图像会出现色偏、模糊和低对比度。许多增强方法在抑制局部特征的同时恢复全局外观；其他方法在频域操作或依赖辅助正则化，但缺乏明确的机制来融合全局和局部证据。我们介绍了GLUE-GAN，这是一种生成式对抗网络，通过显式跨尺度，跨域融合在全局局部统一增强。该模型包括：(1)自适应特征增强编码器，该编码器将空间上下文模块与分组通道自关注结合起来，协同建模不同场景中的空间通道依赖关系；(2)多通道特征聚合增强模块，进行多尺度提取和对齐，统一恢复全局色调和局部纹理；(3)采用小波分解分离低频段和高频频段的全局局部信息增强模块，在全局校正过程中减轻局部偏置。对EUVP、UIEB和UFO-120的评估表明，在色彩保真度、对比度和清晰度方面取得了一致的进展，同时改善了边缘和精细细节的保存。通过将空间信道推理与频率感知处理统一在一个单一的对抗框架中，GLUE-GAN平衡了水下图像增强的全局色彩校正和局部细节保留。

{"title":"GLUE-GAN: Global-local underwater image enhancement generative adversarial network","authors":"Jun Wang , Daoyi Wei , Wentao Shi , Chenyang Liu , Wenlong Jiang , Juan Wang","doi":"10.1016/j.asoc.2026.114686","DOIUrl":"10.1016/j.asoc.2026.114686","url":null,"abstract":"<div><div>Underwater images suffer from color cast, blur, and low contrast due to wavelength-dependent absorption, scattering, and suspended particulates. Many enhancement methods restore global appearance while suppressing local features; others operate in the frequency domain or rely on auxiliary regularization but lack an explicit mechanism to fuse global and local evidence. We introduce GLUE-GAN, a generative adversarial network that globally–locally unifies enhancement through explicit cross-scale, cross-domain fusion. The model comprises: (1) an adaptive feature enhanced encoder that marries spatial context modules with grouped channel-wise self-attention to collaboratively model spatial–channel dependencies across diverse scenes; (2) a multichannel feature aggregation enhancement module that performs multi-scale extraction and alignment to uniformly recover global tone and local textures; and (3) a global–local information enhancement module that uses wavelet decomposition to separate low- and high-frequency bands, processing that mitigates local bias during global correction. Evaluations on EUVP, UIEB, and UFO-120 demonstrate consistent gains in color fidelity, contrast, and sharpness, with improved preservation of edges and fine details. By unifying spatial–channel reasoning with frequency-aware processing in a single adversarial framework, GLUE-GAN balances global color correction and local detail preservation for underwater image enhancement.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114686"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080303","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

HPPRNet: Human parsing enabled pose refinement network for human activity recognition hprnet：用于人类活动识别的支持人类解析的姿态优化网络

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.2026.114700

Vivek Tiwari , Aditi Verma , Mayank Lovanshi

Human activity recognition has attracted much attention recently due to its impressive ability to recognise complex activities. This study aims to perform human activity recognition at a coarse-grain level by employing a human-centric visual analysis method called human parsing. However, to retain the temporal aspects of the video sequences, the proposed design incorporates an integrated human parsing and pose network, which mutually enhance each other. This paper introduces a comprehensive architecture called the Human Parsing-Enabled Pose Refinement Network (HPPRNet), designed to recognise human actions. This study also introduces a novel dataset called SP-VIP (Single Person-Video in Person), which consists of 246 short videos featuring single individuals. This dataset is designed to support all three objectives: parsing, pose estimation, and action recognition. Additionally, the sequential data produced by the video and its refined version are fed into an LSTM network for human activity classification. This approach achieves an accuracy of 94.61% for four activity categories and 92.14% for six categories. Also, the pose-enabled human parsing network produced an overall pixel accuracy of 88.50%, a mean accuracy of 60.50% and a mean IoU of 49.30% while parsing the video frames.

人类活动识别由于其令人印象深刻的识别复杂活动的能力，近年来引起了人们的广泛关注。本研究旨在通过采用以人为中心的视觉分析方法，即人类解析，在粗粒度水平上进行人类活动识别。然而，为了保留视频序列的时间方面，提出的设计结合了一个集成的人类解析和姿态网络，它们相互增强。本文介绍了一种名为Human parsin - enabled Pose Refinement Network （hprnet）的综合架构，用于识别人类的动作。本研究还引入了一个名为SP-VIP（单人视频）的新数据集，该数据集由246个以单人为主角的短视频组成。该数据集旨在支持所有三个目标：解析，姿态估计和动作识别。此外，视频产生的序列数据及其精炼版本被馈送到LSTM网络中用于人类活动分类。该方法对四类活动的准确率为94.61%，对六类活动的准确率为92.14%。此外，在分析视频帧时，支持姿态的人工解析网络产生的总像素精度为88.50%，平均精度为60.50%，平均IoU为49.30%。

{"title":"HPPRNet: Human parsing enabled pose refinement network for human activity recognition","authors":"Vivek Tiwari , Aditi Verma , Mayank Lovanshi","doi":"10.1016/j.asoc.2026.114700","DOIUrl":"10.1016/j.asoc.2026.114700","url":null,"abstract":"<div><div>Human activity recognition has attracted much attention recently due to its impressive ability to recognise complex activities. This study aims to perform human activity recognition at a coarse-grain level by employing a human-centric visual analysis method called human parsing. However, to retain the temporal aspects of the video sequences, the proposed design incorporates an integrated human parsing and pose network, which mutually enhance each other. This paper introduces a comprehensive architecture called the Human Parsing-Enabled Pose Refinement Network (HPPRNet), designed to recognise human actions. This study also introduces a novel dataset called SP-VIP (Single Person-Video in Person), which consists of 246 short videos featuring single individuals. This dataset is designed to support all three objectives: parsing, pose estimation, and action recognition. Additionally, the sequential data produced by the video and its refined version are fed into an LSTM network for human activity classification. This approach achieves an accuracy of 94.61% for four activity categories and 92.14% for six categories. Also, the pose-enabled human parsing network produced an overall pixel accuracy of 88.50%, a mean accuracy of 60.50% and a mean IoU of 49.30% while parsing the video frames.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114700"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080203","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

Frequency domain iterative clustering for boundary-preserving superpixel segmentation 保持边界的超像素分割的频域迭代聚类

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.2026.114717

Junbin Zhuang , Kun Wang , Zhe Yuan , Yunyi Yan

Superpixel segmentation has become a powerful tool in the rapidly developing field of image processing, particularly for subject recognition and object detection. However, spatial superpixel segmentation does not adaptively act according to fine details it may result in over-sharpened features; The absence of accurate quantitative analysis of information loss following segmentation. We observed that different frequency information exerts varying influences on the segmentation outcomes during superpixel segmentation. Low-frequency information has a relatively minor positive effect, whereas diagonal high-frequency information exhibits a greater impact compared to horizontal and vertical high-frequency information. This phenomenon can be attributed to the ability of high-frequency components to capture more boundary and detail information in complex scenes, which is crucial for generating superpixels with clear boundaries and rich semantic information. Based on this observation, Frequency Driven Iterative Clustering (FDIC) technology integrates frequency domain analysis into the superpixel segmentation algorithm which could enable direct evaluation of regional complexity with better generation of initial seeds, and as a result it could achieve high-quality segmentation. Additionally, we introduce a novel no-reference evaluation metric, the Superpixel Peak Signal-to-Noise Ratio (SSNR), which quantifies the error between the original image and the distorted image, providing a more intuitive superpixel quality evaluation metric. Experimental results on the BSDS500 and NYUv2 datasets demonstrate that FDIC achieves superior performance in superpixel segmentation.

在快速发展的图像处理领域，超像素分割已成为一种强有力的工具，特别是在主题识别和目标检测方面。但是，空间超像素分割不能根据细节进行自适应处理，可能导致特征过锐化；对分割后的信息损失缺乏准确的定量分析。我们观察到不同的频率信息对超像素分割结果的影响是不同的。低频信息的正向影响相对较小，而对角线高频信息的正向影响大于水平和垂直高频信息。这种现象可以归因于高频成分能够在复杂场景中捕获更多的边界和细节信息，这对于生成具有清晰边界和丰富语义信息的超像素至关重要。基于此，频率驱动迭代聚类（Frequency Driven Iterative Clustering， FDIC）技术将频域分析集成到超像素分割算法中，可以更好地生成初始种子，直接评估区域复杂度，从而实现高质量的分割。此外，我们引入了一种新的无参考评价指标——超像素峰值信噪比（SSNR），它量化了原始图像与失真图像之间的误差，提供了一种更直观的超像素质量评价指标。在BSDS500和NYUv2数据集上的实验结果表明，FDIC在超像素分割方面取得了优异的性能。

{"title":"Frequency domain iterative clustering for boundary-preserving superpixel segmentation","authors":"Junbin Zhuang , Kun Wang , Zhe Yuan , Yunyi Yan","doi":"10.1016/j.asoc.2026.114717","DOIUrl":"10.1016/j.asoc.2026.114717","url":null,"abstract":"<div><div>Superpixel segmentation has become a powerful tool in the rapidly developing field of image processing, particularly for subject recognition and object detection. However, spatial superpixel segmentation does not adaptively act according to fine details it may result in over-sharpened features; The absence of accurate quantitative analysis of information loss following segmentation. We observed that different frequency information exerts varying influences on the segmentation outcomes during superpixel segmentation. Low-frequency information has a relatively minor positive effect, whereas diagonal high-frequency information exhibits a greater impact compared to horizontal and vertical high-frequency information. This phenomenon can be attributed to the ability of high-frequency components to capture more boundary and detail information in complex scenes, which is crucial for generating superpixels with clear boundaries and rich semantic information. Based on this observation, Frequency Driven Iterative Clustering (FDIC) technology integrates frequency domain analysis into the superpixel segmentation algorithm which could enable direct evaluation of regional complexity with better generation of initial seeds, and as a result it could achieve high-quality segmentation. Additionally, we introduce a novel no-reference evaluation metric, the Superpixel Peak Signal-to-Noise Ratio (SSNR), which quantifies the error between the original image and the distorted image, providing a more intuitive superpixel quality evaluation metric. Experimental results on the BSDS500 and NYUv2 datasets demonstrate that FDIC achieves superior performance in superpixel segmentation.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114717"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080205","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 ant colony optimization method for the mixed storage and pre-marshalling problem considering pre-processing 考虑预处理的混合存储和预编组问题的蚁群优化方法

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.2026.114718

Lebao Wu , Zuhua Jiang , Daofang Chang

This study investigates storage operations in indoor steel plate yards within the shipbuilding industry, focusing on sequential storage following pre-processing. Pre-marshalling moves are permitted during pre-processing intervals to optimize storage configurations. The problem requires determining both storage locations and pre-marshalling moves within a constrained time window to minimize outbound relocations during plate retrieval. To address this, the quantitative relationship between pre-processing time and the number of allowable pre-marshalling moves is analyzed, and a mixed integer programming model is formulated. Key factors affecting outbound relocations are identified, and an estimation formula is developed to quantify the impact of each move on outbound relocations. A novel ant colony optimization approach is proposed, utilizing the estimation formula for heuristic value and storage move selection, and incorporating two strategies to accelerate computation. The proposed method’s effectiveness is validated by comparing it with benchmark approaches from the literature on real-sized instances. Additionally, the effects of two key problem parameters are analyzed.

本研究调查船舶工业室内钢板堆场的储存作业，重点是预处理后的顺序储存。允许在预处理间隔期间进行预编组移动，以优化存储配置。该问题需要在受限的时间窗口内确定存储位置和预编组移动，以最大限度地减少板检索期间的出站重定位。为了解决这个问题，分析了预处理时间与允许的预编组次数之间的定量关系，并建立了一个混合整数规划模型。确定了影响出站搬迁的关键因素，并开发了一个估计公式来量化每次搬迁对出站搬迁的影响。提出了一种新的蚁群优化方法，利用启发式值和存储移动选择的估计公式，结合两种策略来加速计算。通过与文献中的基准方法在实际实例上的比较，验证了该方法的有效性。此外，还分析了两个关键问题参数的影响。

{"title":"An ant colony optimization method for the mixed storage and pre-marshalling problem considering pre-processing","authors":"Lebao Wu , Zuhua Jiang , Daofang Chang","doi":"10.1016/j.asoc.2026.114718","DOIUrl":"10.1016/j.asoc.2026.114718","url":null,"abstract":"<div><div>This study investigates storage operations in indoor steel plate yards within the shipbuilding industry, focusing on sequential storage following pre-processing. Pre-marshalling moves are permitted during pre-processing intervals to optimize storage configurations. The problem requires determining both storage locations and pre-marshalling moves within a constrained time window to minimize outbound relocations during plate retrieval. To address this, the quantitative relationship between pre-processing time and the number of allowable pre-marshalling moves is analyzed, and a mixed integer programming model is formulated. Key factors affecting outbound relocations are identified, and an estimation formula is developed to quantify the impact of each move on outbound relocations. A novel ant colony optimization approach is proposed, utilizing the estimation formula for heuristic value and storage move selection, and incorporating two strategies to accelerate computation. The proposed method’s effectiveness is validated by comparing it with benchmark approaches from the literature on real-sized instances. Additionally, the effects of two key problem parameters are analyzed.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114718"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080304","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

From rewards to decisions: State-aware multi-objective group decision-making reinforcement learning for volatile futures 从奖励到决策：波动期货的状态感知多目标群体决策强化学习

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.2026.114719

Qing Zhu , Baoqin Xie , Shan Liu , Yuze Li

Amid global geopolitical turbulence and increasing financial market volatility, futures trading strategies face significant challenges in dynamically balancing risk, cost, and return. Existing studies predominantly rely on static reward functions or single reinforcement learning algorithms, limiting their adaptability to sudden market shifts and strategy degradation. This study proposes a multi-agent reinforcement learning framework (MA-D-RCR) that integrates dynamic reward functions, group decision-making, and continual learning to overcome the adaptability limitations of traditional futures trading strategies. The core contribution lies in the synergistic co-evolution and closed-loop feedback among these mechanisms, enabling trading strategies to effectively adjust to dynamic market conditions while supporting knowledge transfer and accumulation. Empirical analysis based on gold futures data from May 30, 2022, to May 26, 2025, demonstrates that the proposed framework exhibits strong adaptability and stability in volatile markets, achieving an annualized return of 68.83% and a Sharpe ratio of 6.83, significantly outperforming static single-objective baselines. The framework effectively coordinates multiple policies, balances multiple objectives, and maintains robust performance across different market conditions. This research provides a novel approach for intelligent financial decision-making, emphasizing environment-aware strategy adaptation, collaborative optimization across agents, dynamic multi-objective trade-offs, and continual learning capabilities, thereby offering valuable insights for trading strategies in highly volatile markets.

在全球地缘政治动荡和金融市场波动加剧的背景下，期货交易策略在动态平衡风险、成本和回报方面面临着重大挑战。现有的研究主要依赖于静态奖励函数或单一的强化学习算法，限制了它们对市场突然变化和策略退化的适应性。为了克服传统期货交易策略的适应性限制，本文提出了一种集动态奖励函数、群体决策和持续学习于一体的多智能体强化学习框架。核心贡献在于这些机制之间的协同进化和闭环反馈，使交易策略能够有效地适应动态市场条件，同时支持知识的转移和积累。基于2022年5月30日至2025年5月26日黄金期货数据的实证分析表明，该框架在波动市场中表现出较强的适应性和稳定性，年化收益率为68.83%，夏普比率为6.83，显著优于静态单目标基准。该框架有效地协调多个政策，平衡多个目标，并在不同的市场条件下保持强劲的表现。该研究为智能金融决策提供了一种新颖的方法，强调环境感知策略适应，跨代理协作优化，动态多目标权衡和持续学习能力，从而为高度波动的市场中的交易策略提供了有价值的见解。

{"title":"From rewards to decisions: State-aware multi-objective group decision-making reinforcement learning for volatile futures","authors":"Qing Zhu , Baoqin Xie , Shan Liu , Yuze Li","doi":"10.1016/j.asoc.2026.114719","DOIUrl":"10.1016/j.asoc.2026.114719","url":null,"abstract":"<div><div>Amid global geopolitical turbulence and increasing financial market volatility, futures trading strategies face significant challenges in dynamically balancing risk, cost, and return. Existing studies predominantly rely on static reward functions or single reinforcement learning algorithms, limiting their adaptability to sudden market shifts and strategy degradation. This study proposes a multi-agent reinforcement learning framework (MA-D-RCR) that integrates dynamic reward functions, group decision-making, and continual learning to overcome the adaptability limitations of traditional futures trading strategies. The core contribution lies in the synergistic co-evolution and closed-loop feedback among these mechanisms, enabling trading strategies to effectively adjust to dynamic market conditions while supporting knowledge transfer and accumulation. Empirical analysis based on gold futures data from May 30, 2022, to May 26, 2025, demonstrates that the proposed framework exhibits strong adaptability and stability in volatile markets, achieving an annualized return of 68.83% and a Sharpe ratio of 6.83, significantly outperforming static single-objective baselines. The framework effectively coordinates multiple policies, balances multiple objectives, and maintains robust performance across different market conditions. This research provides a novel approach for intelligent financial decision-making, emphasizing environment-aware strategy adaptation, collaborative optimization across agents, dynamic multi-objective trade-offs, and continual learning capabilities, thereby offering valuable insights for trading strategies in highly volatile markets.</div></div>","PeriodicalId":50737,"journal":{"name":"Applied Soft Computing","volume":"191 ","pages":"Article 114719"},"PeriodicalIF":6.6,"publicationDate":"2026-01-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146080204","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