Scholarlink

<p>Support Vector Machines (SVM) have been widely used in supervised learning due to their strong generalization capabilities. SVM + extends this framework under the Learning Using Privileged Information (LUPI) paradigm, incorporating additional training-stage information unavailable during inference. However, SVM + involves the tuning of four hyperparameters, doubling the dimensionality of the optimization problem compared to standard SVM, and thereby increasing computational cost substantially. To address this challenge, this work introduces GAGS (Genetic Algorithm-Based Grid Search), a methodological hybridization framework tailored for SVM + model selection that integrates grid-based structured search, adaptive precision through logarithmic mapping, and advanced parallelization strategies. The term 'Grid Search' in GAGS refers to the grid-based structural partitioning used for initial exploration and parallel computation, and not to a traditional exhaustive search. By distributing computations across independent processing units, GAGS enables structured exploration of the hyperparameter space and reduces the likelihood of premature convergence and local optima entrapment, while substantially reducing execution time under the evaluated settings. The proposed GAGS-SVM + framework is evaluated on a binary classification task using the MNIST and CWRU datasets under a rigorous protocol of 20 independent runs. Experimental results show that the GAGS methodology achieves statistically higher mean accuracy than the tested baselines under a fixed evaluation budget, reducing error rates in the vision domain test (MNIST) to below 4.5% and consistently matching or surpassing the metaheuristic baselines on both datasets under an equal evaluation budget and identical data splits across methods. The performance gains are further reflected in reduced test errors under data scarcity scenarios within the evaluated datasets. Furthermore, the parallel implementation of GAGS achieves up to ≈4 × speed-up under the reported hardware configuration, supporting its scalability within the evaluated computational regime and its suitability for resource-intensive kernel-based optimization tasks characterized by high-cost QP solves.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

GAGS: parallelized genetic algorithm-based grid search for SVM + model optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月14日

查看全文
  • J. D. Fernández,
  • J. Vega,
  • S. Dormido-Canto

摘要

Support Vector Machines (SVM) have been widely used in supervised learning due to their strong generalization capabilities. SVM + extends this framework under the Learning Using Privileged Information (LUPI) paradigm, incorporating additional training-stage information unavailable during inference. However, SVM + involves the tuning of four hyperparameters, doubling the dimensionality of the optimization problem compared to standard SVM, and thereby increasing computational cost substantially. To address this challenge, this work introduces GAGS (Genetic Algorithm-Based Grid Search), a methodological hybridization framework tailored for SVM + model selection that integrates grid-based structured search, adaptive precision through logarithmic mapping, and advanced parallelization strategies. The term 'Grid Search' in GAGS refers to the grid-based structural partitioning used for initial exploration and parallel computation, and not to a traditional exhaustive search. By distributing computations across independent processing units, GAGS enables structured exploration of the hyperparameter space and reduces the likelihood of premature convergence and local optima entrapment, while substantially reducing execution time under the evaluated settings. The proposed GAGS-SVM + framework is evaluated on a binary classification task using the MNIST and CWRU datasets under a rigorous protocol of 20 independent runs. Experimental results show that the GAGS methodology achieves statistically higher mean accuracy than the tested baselines under a fixed evaluation budget, reducing error rates in the vision domain test (MNIST) to below 4.5% and consistently matching or surpassing the metaheuristic baselines on both datasets under an equal evaluation budget and identical data splits across methods. The performance gains are further reflected in reduced test errors under data scarcity scenarios within the evaluated datasets. Furthermore, the parallel implementation of GAGS achieves up to ≈4 × speed-up under the reported hardware configuration, supporting its scalability within the evaluated computational regime and its suitability for resource-intensive kernel-based optimization tasks characterized by high-cost QP solves.