Pub Date : 2025-12-12DOI: 10.1016/j.swevo.2025.102233
Kamran Dashti Maljaei, S. Kamal Chaharsooghi, Ali Husseinzadeh Kashan
A significant gap persists between theoretical scheduling models and the complex realities of industrial manufacturing. This study bridges this gap by proposing a comprehensive yet tractable framework for jointly scheduling production and condition-based preventive maintenance (PM) in a three-stage hybrid flow shop. We introduce a novel mixed-integer linear programming (MILP) model that integrates a suite of realistic constraints, using machine performance degradation as an endogenous trigger for PM activities. To solve this NP-hard problem, we apply Dynamic Programming (DP) for exact validation on small instances and develop seven meta-heuristics, including four novel hybrid strategies, for large-scale applications. The framework is validated through extensive computational experiments and a real-world automotive case study, with results indicating that the proposed HSGA-I algorithm delivers a superior trade-off between solution quality and computational efficiency. Ultimately, the framework provides a practical decision-support tool for managers, enabling tangible improvements such as an estimated 9% reduction in operational costs and a 15% decrease in machine downtime, by optimizing the trade-off between short-term production targets and long-term machine reliability.
{"title":"Optimizing three-stage hybrid flow shop scheduling: A dynamic programming and hybrid meta-heuristic framework for joint production and preventive maintenance under real-world constraints","authors":"Kamran Dashti Maljaei, S. Kamal Chaharsooghi, Ali Husseinzadeh Kashan","doi":"10.1016/j.swevo.2025.102233","DOIUrl":"10.1016/j.swevo.2025.102233","url":null,"abstract":"<div><div>A significant gap persists between theoretical scheduling models and the complex realities of industrial manufacturing. This study bridges this gap by proposing a comprehensive yet tractable framework for jointly scheduling production and condition-based preventive maintenance (PM) in a three-stage hybrid flow shop. We introduce a novel mixed-integer linear programming (MILP) model that integrates a suite of realistic constraints, using machine performance degradation as an endogenous trigger for PM activities. To solve this NP-hard problem, we apply Dynamic Programming (DP) for exact validation on small instances and develop seven meta-heuristics, including four novel hybrid strategies, for large-scale applications. The framework is validated through extensive computational experiments and a real-world automotive case study, with results indicating that the proposed HSGA-I algorithm delivers a superior trade-off between solution quality and computational efficiency. Ultimately, the framework provides a practical decision-support tool for managers, enabling tangible improvements such as an estimated 9% reduction in operational costs and a 15% decrease in machine downtime, by optimizing the trade-off between short-term production targets and long-term machine reliability.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102233"},"PeriodicalIF":8.5,"publicationDate":"2025-12-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145736998","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}
Pub Date : 2025-12-11DOI: 10.1016/j.swevo.2025.102250
Alper Hamzadayı , Mehmet Ali Arvas
The identical parallel machine scheduling problem with a single server and sequence-dependent setup times is a challenging optimization problem with important applications in manufacturing and service industries. In such environments, several machines depend on a common server to perform setup operations before production can begin, which creates strong interdependencies and demands more effective scheduling strategies. This characteristic highlights the practical relevance of the problem. The interaction between machine availability and server operations often becomes a critical bottleneck. This study introduces two complementary approaches. The first is an exact method based on a novel arc-based mixed-integer linear programming (ABF) model, which extends the modeling capability of existing formulations by capturing server-related constraints more effectively. The second is an approximation method built on an Iterated Greedy (IG) algorithm. The IG procedure is improved by two evaluation mechanisms: one model-based evaluation derived from the proposed ABF model, and another employing a greedy randomized adaptive search procedure (GRASP)-based strategy that integrates greedy selection, randomization, and reconstruction to enhance solution quality. Computational experiments are conducted on existing benchmark instances. The results show that the proposed ABF model performs well on small and medium-sized instances compared to existing exact methods, while the IG variants, particularly the proposed GRASP-based version, deliver strong performance against state-of-the-art metaheuristics developed for this problem. In addition, 21 new best-known solutions are reported, further demonstrating the effectiveness of the proposed approaches.
{"title":"Arc-based formulation and GRASP-enhanced iterated greedy algorithm for identical parallel machine scheduling with a common server","authors":"Alper Hamzadayı , Mehmet Ali Arvas","doi":"10.1016/j.swevo.2025.102250","DOIUrl":"10.1016/j.swevo.2025.102250","url":null,"abstract":"<div><div>The identical parallel machine scheduling problem with a single server and sequence-dependent setup times is a challenging optimization problem with important applications in manufacturing and service industries. In such environments, several machines depend on a common server to perform setup operations before production can begin, which creates strong interdependencies and demands more effective scheduling strategies. This characteristic highlights the practical relevance of the problem. The interaction between machine availability and server operations often becomes a critical bottleneck. This study introduces two complementary approaches. The first is an exact method based on a novel arc-based mixed-integer linear programming (ABF) model, which extends the modeling capability of existing formulations by capturing server-related constraints more effectively. The second is an approximation method built on an Iterated Greedy (IG) algorithm. The IG procedure is improved by two evaluation mechanisms: one model-based evaluation derived from the proposed ABF model, and another employing a greedy randomized adaptive search procedure (GRASP)-based strategy that integrates greedy selection, randomization, and reconstruction to enhance solution quality. Computational experiments are conducted on existing benchmark instances. The results show that the proposed ABF model performs well on small and medium-sized instances compared to existing exact methods, while the IG variants, particularly the proposed GRASP-based version, deliver strong performance against state-of-the-art metaheuristics developed for this problem. In addition, 21 new best-known solutions are reported, further demonstrating the effectiveness of the proposed approaches.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102250"},"PeriodicalIF":8.5,"publicationDate":"2025-12-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145736995","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}
Pub Date : 2025-12-11DOI: 10.1016/j.swevo.2025.102251
Shihong Yin, Zhengrong Xiang
This paper addresses the multi-objective collaborative path planning problem for heterogeneous autonomous underwater vehicles (AUVs) in complex underwater environments. The problem involves multiple conflicting optimization objectives, such as task collaboration time, risk cost, and energy consumption, while also facing constraints imposed by complex environmental factors, such as ocean current interference and cluttered threat regions. To tackle this challenge, a novel adaptive dual-ranking constrained multi-objective differential evolution (DR-CMODE) algorithm is proposed. This algorithm combines Pareto dominance ranking and constraint dominance ranking mechanisms, adaptively adjusting the weights between them to balance objective optimization and constraint satisfaction. Additionally, the algorithm integrates four differential evolution operators to enhance solution diversity and convergence efficiency. Extensive numerical simulations demonstrate that DR-CMODE can effectively generate feasible and high-quality AUV paths in cluttered underwater environments, achieving an optimal trade-off among efficiency, safety, and energy consumption. To further verify the robustness of the algorithm, the DR-CMODE is applied to solve the AUV collaborative path planning problem in complex simulated environments that incorporate time-varying ocean currents and noise disturbances. It significantly outperforms eleven advanced constrained multi-objective optimization methods in terms of Hypervolume metrics, solution robustness, and convergence speed. The source code and data are available at https://github.com/Shihong-Yin/DR-CMODE-MOCP_AUV.
{"title":"Multi-objective collaborative path planning for heterogeneous autonomous underwater vehicles in cluttered environments","authors":"Shihong Yin, Zhengrong Xiang","doi":"10.1016/j.swevo.2025.102251","DOIUrl":"10.1016/j.swevo.2025.102251","url":null,"abstract":"<div><div>This paper addresses the multi-objective collaborative path planning problem for heterogeneous autonomous underwater vehicles (AUVs) in complex underwater environments. The problem involves multiple conflicting optimization objectives, such as task collaboration time, risk cost, and energy consumption, while also facing constraints imposed by complex environmental factors, such as ocean current interference and cluttered threat regions. To tackle this challenge, a novel adaptive dual-ranking constrained multi-objective differential evolution (DR-CMODE) algorithm is proposed. This algorithm combines Pareto dominance ranking and constraint dominance ranking mechanisms, adaptively adjusting the weights between them to balance objective optimization and constraint satisfaction. Additionally, the algorithm integrates four differential evolution operators to enhance solution diversity and convergence efficiency. Extensive numerical simulations demonstrate that DR-CMODE can effectively generate feasible and high-quality AUV paths in cluttered underwater environments, achieving an optimal trade-off among efficiency, safety, and energy consumption. To further verify the robustness of the algorithm, the DR-CMODE is applied to solve the AUV collaborative path planning problem in complex simulated environments that incorporate time-varying ocean currents and noise disturbances. It significantly outperforms eleven advanced constrained multi-objective optimization methods in terms of Hypervolume metrics, solution robustness, and convergence speed. The source code and data are available at <span><span>https://github.com/Shihong-Yin/DR-CMODE-MOCP_AUV</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102251"},"PeriodicalIF":8.5,"publicationDate":"2025-12-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145736997","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}
Pub Date : 2025-12-10DOI: 10.1016/j.swevo.2025.102254
Michał K. Tomczyk, Miłosz Kadziński
We propose a preference-learning algorithm tailored for interactive evolutionary multi-objective optimization. The method estimates the parameters of an assumed preference model from incomplete feedback provided by the decision maker (DM), addressing two challenges: (i) identifying compatible model instances even when preference information strongly constrains the parameter space, and (ii) generating a diverse, approximately uniform set of models to support robust decision making. These goals are achieved via an evolutionary process that iteratively refines a population of models using specialized operators. The algorithm prioritizes models that are both compatible with the elicited preferences and sufficiently dissimilar from their nearest neighbors, thereby promoting a well-distributed coverage of the feasible parameter space. To improve computational efficiency, we introduce a queue-based mechanism that directs the evolutionary process with minimal overhead, enhancing responsiveness for interactive use. We evaluate the proposed method in two complementary settings: first, as a standalone sampler, and second, embedded within an evolutionary multi-objective optimizer to demonstrate its utility for interactive decision support.
{"title":"Efficient preference learning algorithm for interactive evolutionary multi-objective optimization","authors":"Michał K. Tomczyk, Miłosz Kadziński","doi":"10.1016/j.swevo.2025.102254","DOIUrl":"10.1016/j.swevo.2025.102254","url":null,"abstract":"<div><div>We propose a preference-learning algorithm tailored for interactive evolutionary multi-objective optimization. The method estimates the parameters of an assumed preference model from incomplete feedback provided by the decision maker (DM), addressing two challenges: (i) identifying compatible model instances even when preference information strongly constrains the parameter space, and (ii) generating a diverse, approximately uniform set of models to support robust decision making. These goals are achieved via an evolutionary process that iteratively refines a population of models using specialized operators. The algorithm prioritizes models that are both compatible with the elicited preferences and sufficiently dissimilar from their nearest neighbors, thereby promoting a well-distributed coverage of the feasible parameter space. To improve computational efficiency, we introduce a queue-based mechanism that directs the evolutionary process with minimal overhead, enhancing responsiveness for interactive use. We evaluate the proposed method in two complementary settings: first, as a standalone sampler, and second, embedded within an evolutionary multi-objective optimizer to demonstrate its utility for interactive decision support.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102254"},"PeriodicalIF":8.5,"publicationDate":"2025-12-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145736994","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}
Pub Date : 2025-12-08DOI: 10.1016/j.swevo.2025.102237
Zhenxing Zhang , Tianxian Zhang , Xiangliang Xu , Zicheng Wang , Lingjiang Kong , Kaibo Shi , Witold Pedrycz
This paper presents a complementary heterogeneity-driven adaptive balance search method for cognitive-only PSOs, designed to overcome the limitation of mixed roles (exploration and exploitation) of particles in traditional cognitive-only PSOs. The proposed method enables fine-grained control over the ratio of particles with distinct roles and can be seamlessly incorporated into various cognitive-only PSO variants. Specifically, the proposed method includes: (1) A generalized complementary heterogeneous PSO framework, which consists of two nearly heterogeneous update channels and two independent subswarms. Each channel directs its corresponding subswarm to specialize in either exploration or exploitation, thereby mitigating performance degradation caused by mixed roles in traditional cognitive-only PSOs. Furthermore, in a manner analogous to classical cognitive-only PSOs, we redefine several key terms to facilitate the seamless integration of diverse cognitive-only PSOs. (2) An adaptive balance search strategy, which dynamically selects particles for each iteration. This strategy achieves precise, stage-aware control over the particle ratio while preserving role specialization, thus enhancing the ability of traditional cognitive-only PSOs to balance exploration and exploitation. Extensive experiments verify the generalization and significant performance improvements delivered by the proposed method.
{"title":"A complementary heterogeneity-driven adaptive balance search method for cognitive-only particle swarm optimization family","authors":"Zhenxing Zhang , Tianxian Zhang , Xiangliang Xu , Zicheng Wang , Lingjiang Kong , Kaibo Shi , Witold Pedrycz","doi":"10.1016/j.swevo.2025.102237","DOIUrl":"10.1016/j.swevo.2025.102237","url":null,"abstract":"<div><div>This paper presents a complementary heterogeneity-driven adaptive balance search method for cognitive-only PSOs, designed to overcome the limitation of mixed roles (exploration and exploitation) of particles in traditional cognitive-only PSOs. The proposed method enables fine-grained control over the ratio of particles with distinct roles and can be seamlessly incorporated into various cognitive-only PSO variants. Specifically, the proposed method includes: (1) A generalized complementary heterogeneous PSO framework, which consists of two nearly heterogeneous update channels and two independent subswarms. Each channel directs its corresponding subswarm to specialize in either exploration or exploitation, thereby mitigating performance degradation caused by mixed roles in traditional cognitive-only PSOs. Furthermore, in a manner analogous to classical cognitive-only PSOs, we redefine several key terms to facilitate the seamless integration of diverse cognitive-only PSOs. (2) An adaptive balance search strategy, which dynamically selects particles for each iteration. This strategy achieves precise, stage-aware control over the particle ratio while preserving role specialization, thus enhancing the ability of traditional cognitive-only PSOs to balance exploration and exploitation. Extensive experiments verify the generalization and significant performance improvements delivered by the proposed method.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102237"},"PeriodicalIF":8.5,"publicationDate":"2025-12-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145736993","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}
Pub Date : 2025-12-07DOI: 10.1016/j.swevo.2025.102246
Andrejaana Andova, Jordan N. Cork, Tea Tušar, Bogdan Filipič
Algorithm selection in optimization is often done by considering a single best-performing algorithm per problem. However, sometimes multiple algorithms perform comparably well on the same optimization problem, and in such cases, it would be appropriate to consider all of them as best performing. Hence, this work proposes an algorithm selection methodology that enables the identification and prediction of multiple algorithms as best performing. More specifically, the methodology involves first identifying the best-performing algorithms using statistical tests that show when the algorithms perform comparably well. Then, these algorithms are set as targets to machine learning models that can predict multiple algorithms as best performing. Finally, an evaluation measure is introduced to assess the performance of the algorithm selection models. The proposed methodology is applied to constrained multiobjective optimization.
{"title":"A methodology for multi-label algorithm selection in constrained multiobjective optimization","authors":"Andrejaana Andova, Jordan N. Cork, Tea Tušar, Bogdan Filipič","doi":"10.1016/j.swevo.2025.102246","DOIUrl":"10.1016/j.swevo.2025.102246","url":null,"abstract":"<div><div>Algorithm selection in optimization is often done by considering a single best-performing algorithm per problem. However, sometimes multiple algorithms perform comparably well on the same optimization problem, and in such cases, it would be appropriate to consider all of them as best performing. Hence, this work proposes an algorithm selection methodology that enables the identification and prediction of multiple algorithms as best performing. More specifically, the methodology involves first identifying the best-performing algorithms using statistical tests that show when the algorithms perform comparably well. Then, these algorithms are set as targets to machine learning models that can predict multiple algorithms as best performing. Finally, an evaluation measure is introduced to assess the performance of the algorithm selection models. The proposed methodology is applied to constrained multiobjective optimization.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102246"},"PeriodicalIF":8.5,"publicationDate":"2025-12-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145736942","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}
Pub Date : 2025-12-06DOI: 10.1016/j.swevo.2025.102245
A.K. Bhurjee , Mridul Patel , P. Kumar
A complex problem involving multiple objectives in fractional optimization, in which the coefficients of both objectives and constraints are expressed as intervals, is investigated in this study. An extended framework based on a generalized interval vector space is proposed, through which a linear transformation to real space is established, allowing a meaningful comparison between interval parameters. By means of this transformation, the interval fractional optimization problem is reformulated into a conventional multiobjective optimization problem. The existence and characterization of efficient solutions for the multiobjective interval problem are analyzed. To validate the proposed approach, numerical experiments and an inventory model under uncertain demand and holding costs have been presented. The model has been solved using several metaheuristic algorithms, Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Grey Wolf Optimizer (GWO), Harris Hawks Optimization (HHO), and MOEA/D-DE, and the results have been compared with analytical benchmarks. The proposed framework has achieved consistent and stable profit and cost intervals across all algorithms, while the Friedman test has indicated that MOEA/D-DE outperforms others with the most balanced performance. Scalability tests on higher-dimensional problems have further demonstrated the robustness and practical applicability of the proposed method.
{"title":"An extended interval vector space for the multiobjective fractional optimization problem with application to the inventory model","authors":"A.K. Bhurjee , Mridul Patel , P. Kumar","doi":"10.1016/j.swevo.2025.102245","DOIUrl":"10.1016/j.swevo.2025.102245","url":null,"abstract":"<div><div>A complex problem involving multiple objectives in fractional optimization, in which the coefficients of both objectives and constraints are expressed as intervals, is investigated in this study. An extended framework based on a generalized interval vector space is proposed, through which a linear transformation to real space is established, allowing a meaningful comparison between interval parameters. By means of this transformation, the interval fractional optimization problem is reformulated into a conventional multiobjective optimization problem. The existence and characterization of efficient solutions for the multiobjective interval problem are analyzed. To validate the proposed approach, numerical experiments and an inventory model under uncertain demand and holding costs have been presented. The model has been solved using several metaheuristic algorithms, Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Grey Wolf Optimizer (GWO), Harris Hawks Optimization (HHO), and MOEA/D-DE, and the results have been compared with analytical benchmarks. The proposed framework has achieved consistent and stable profit and cost intervals across all algorithms, while the Friedman test has indicated that MOEA/D-DE outperforms others with the most balanced performance. Scalability tests on higher-dimensional problems have further demonstrated the robustness and practical applicability of the proposed method.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102245"},"PeriodicalIF":8.5,"publicationDate":"2025-12-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145736992","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}
Vaccine hesitancy, magnified during the COVID-19 pandemic, poses a major challenge to global health. While social media offers a real-time window into public sentiment, extracting actionable insights remains difficult. We introduce PULSE (PSO-Utilized LLM-based Stance Exploration), a novel three-stage framework that uniquely integrates optimization, summarization, and explainability for analyzing vaccine-hesitant discourse. In Stage 1, we perform multi-label classification of tweets into 12 overlapping categories of hesitancy using a Nested Long Short-Term Memory (LSTM) architecture, enhanced with attention layers and Particle Swarm Optimization (PSO)—the first application of PSO in this context. We further address class imbalance through ConceptNet-based data augmentation and employ Twitter-specific XLM-RoBERTa embeddings for robust contextual representation. Our best-performing model shows a 23.8% relative accuracy improvement over the baseline, i.e., Nested LSTM, along with higher macro and weighted F1-scores. Stage 2 introduces the first use of LLMs (e.g., GPT-4o, DeepSeek, Gemini) for abstractive summarization of vaccine-hesitant tweets, paired with a dual evaluation strategy using both human and LLM judges—an innovative step toward scalable and high-quality summary validation. Stage 3 enhances transparency via LIME, providing interpretable, token-level rationale behind predictions. To our knowledge, this is the first unified framework combining PSO-driven classification, LLM-based summarization and evaluation, and explainable AI to study vaccine hesitancy on social media. Further, qualitative and quantitative analysis have been performed along with the statistical significance t-test to get the in-depth analysis. Results obtained reveal that our proposed framework shows superior performance over the state-of-the-art methods. The code of this paper is avaiable at https://github.com/anubhavsinghgtm/pulse.
{"title":"PULSE: A Multi-stage Artificial Intelligence Framework for Analyzing Vaccine Hesitancy on Twitter using Particle Swarm Optimization and Large Language Models","authors":"Anubhav Singh , Naveen Saini , Konstantinos Zervoudakis , Vikas Kumar Tiwari","doi":"10.1016/j.swevo.2025.102218","DOIUrl":"10.1016/j.swevo.2025.102218","url":null,"abstract":"<div><div>Vaccine hesitancy, magnified during the COVID-19 pandemic, poses a major challenge to global health. While social media offers a real-time window into public sentiment, extracting actionable insights remains difficult. We introduce <em>PULSE</em> (PSO-Utilized LLM-based Stance Exploration), a novel three-stage framework that uniquely integrates optimization, summarization, and explainability for analyzing vaccine-hesitant discourse. In <em>Stage 1</em>, we perform multi-label classification of tweets into 12 overlapping categories of hesitancy using a <em>Nested Long Short-Term Memory (LSTM)</em> architecture, enhanced with <em>attention layers</em> and <em>Particle Swarm Optimization (PSO)</em>—the first application of PSO in this context. We further address class imbalance through <em>ConceptNet-based data augmentation</em> and employ Twitter-specific XLM-RoBERTa embeddings for robust contextual representation. Our best-performing model shows a <em>23.8% relative accuracy improvement</em> over the baseline, i.e., Nested LSTM, along with higher macro and weighted F1-scores. <em>Stage 2</em> introduces the <em>first use of LLMs</em> (e.g., GPT-4o, DeepSeek, Gemini) for abstractive summarization of vaccine-hesitant tweets, paired with a dual evaluation strategy using both human and LLM judges—an innovative step toward scalable and high-quality summary validation. <em>Stage 3</em> enhances transparency via <em>LIME</em>, providing interpretable, token-level rationale behind predictions. To our knowledge, this is the first unified framework combining PSO-driven classification, LLM-based summarization and evaluation, and explainable AI to study vaccine hesitancy on social media. Further, qualitative and quantitative analysis have been performed along with the statistical significance t-test to get the in-depth analysis. Results obtained reveal that our proposed framework shows superior performance over the state-of-the-art methods. The code of this paper is avaiable at <span><span>https://github.com/anubhavsinghgtm/pulse</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102218"},"PeriodicalIF":8.5,"publicationDate":"2025-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145684854","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}
Pub Date : 2025-12-04DOI: 10.1016/j.swevo.2025.102248
Hao Gao, Junwei Wang, Hongfeng Wang
Designing appropriate search operators is crucial for improving the performance of differential evolution (DE). In conventional DE, the design of operators entails expensive and labor-intensive experiments by experts. The structure of these handcrafted operators is frozen in running, which leads to fixed search patterns in DE and limits its ability to flexibly adapt to the diverse characteristics of the fitness landscape. To solve the above challenges, this paper proposes a concept of dynamically generating mutation operators and further designs a structural fuzzy generative differential evolution (SFG-DE). First, a structural fuzzy generative (SFG) mutation strategy integrates fuzzy logic, Q-learning, and a parameter adaptation mechanism to automatically generate the mathematical model. “Fuzzy generative” refers to a fuzzy logic-driven generation mechanism. Second, an estimation of univariate Gaussian distribution (EUGD) mutation strategy generates mutation vectors based on samples from Gaussian distributions to reduce the attractiveness of the basin to the population. Third, a selection mechanism with metropolis criterion and individual regeneration (MCIR) maintains population diversity by processing solutions that cannot be further improved. The search trajectory network is introduced to explain the SFG-DE behavior, making the algorithmic decision transparent and enhancing user trust. In numerical simulations, SFG-DE achieved average Friedman ranks of 2.42, 3.69, and 2.70 and average Kruskal–Wallis ranks of 104.50, 115.21, and 32.65—ranking 1st, 2nd, and 1st in the eight renowned algorithms, eight winners, and six recent variants, respectively. The results indicate that SFG-DE exhibits highly competitive performance across a broad spectrum of benchmarks and competitors.
{"title":"SFG-DE: An explainable and evolvable differential evolution for learning to generate operator structures","authors":"Hao Gao, Junwei Wang, Hongfeng Wang","doi":"10.1016/j.swevo.2025.102248","DOIUrl":"10.1016/j.swevo.2025.102248","url":null,"abstract":"<div><div>Designing appropriate search operators is crucial for improving the performance of differential evolution (DE). In conventional DE, the design of operators entails expensive and labor-intensive experiments by experts. The structure of these handcrafted operators is frozen in running, which leads to fixed search patterns in DE and limits its ability to flexibly adapt to the diverse characteristics of the fitness landscape. To solve the above challenges, this paper proposes a concept of dynamically generating mutation operators and further designs a structural fuzzy generative differential evolution (SFG-DE). First, a structural fuzzy generative (SFG) mutation strategy integrates fuzzy logic, Q-learning, and a parameter adaptation mechanism to automatically generate the mathematical model. “Fuzzy generative” refers to a fuzzy logic-driven generation mechanism. Second, an estimation of univariate Gaussian distribution (EUGD) mutation strategy generates mutation vectors based on samples from Gaussian distributions to reduce the attractiveness of the basin to the population. Third, a selection mechanism with metropolis criterion and individual regeneration (MCIR) maintains population diversity by processing solutions that cannot be further improved. The search trajectory network is introduced to explain the SFG-DE behavior, making the algorithmic decision transparent and enhancing user trust. In numerical simulations, SFG-DE achieved average Friedman ranks of 2.42, 3.69, and 2.70 and average Kruskal–Wallis ranks of 104.50, 115.21, and 32.65—ranking 1st, 2nd, and 1st in the eight renowned algorithms, eight winners, and six recent variants, respectively. The results indicate that SFG-DE exhibits highly competitive performance across a broad spectrum of benchmarks and competitors.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102248"},"PeriodicalIF":8.5,"publicationDate":"2025-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145684941","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}
This study analyzes the literature and reviews the trends and development trajectories of Particle Swarm Optimization (PSO)-based scheduling. Main Path and Cluster Analysis identify the seminal features introduced to improve PSO, and the major application areas. This serves as the basis for discussing computational advancements. The findings suggest that PSO is most developed in flow-shop scheduling, with its evolution progressing from single- to multi-objective optimization. The main application has shifted from production to advanced computing and energy management, indicating the growing influence of AI, renewables and energy storage. The shift towards mass customization explains the projected growth of flexible job-shop scheduling.
{"title":"Tracing the evolution of Particle Swarm Optimization in scheduling: A systematic review using main path analysis","authors":"Kuo-Ching Ying , Pourya Pourhejazy , Kuan-Lun Huang","doi":"10.1016/j.swevo.2025.102239","DOIUrl":"10.1016/j.swevo.2025.102239","url":null,"abstract":"<div><div>This study analyzes the literature and reviews the trends and development trajectories of Particle Swarm Optimization (PSO)-based scheduling. Main Path and Cluster Analysis identify the seminal features introduced to improve PSO, and the major application areas. This serves as the basis for discussing computational advancements. The findings suggest that PSO is most developed in flow-shop scheduling, with its evolution progressing from single- to multi-objective optimization. The main application has shifted from production to advanced computing and energy management, indicating the growing influence of AI, renewables and energy storage. The shift towards mass customization explains the projected growth of flexible job-shop scheduling.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"100 ","pages":"Article 102239"},"PeriodicalIF":8.5,"publicationDate":"2025-12-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145684855","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}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号