Scholarlink

<p>The Differential Evolution (DE) algorithm is a well-known evolutionary algorithm to deal with global optimization problems. However, it shows the shortcomings of inefficient diversity and slow convergence while dealing with complex problems. In this paper, we propose a modified version of the DE, denoted as DML-DE, by integrating with several search strategies, including population partitioning, a multi-learning strategy based on a new guiding vector, and statistical distributions and successful historical information-based parameter settings. The population partitioning is applied to categorize the individuals based on their fitness so that appropriate search rules can be assigned to them. The multi-learning strategy defines three different mutation rules to ensure sufficient and balanced levels of diversity and convergence speed. The DML-DE is validated on standard CEC2017 problems to verify its efficiency in solving continuous optimization problems, and later, extended to its binary version using a transfer function to find the highly effective feature subsets for the feature selection (FS) problems. A collection of 24 datasets from the UCI repository is used to analyze its performance. The results obtained by DML-DE are compared with 11 metaheuristics using various evaluation measures, which demonstrate that DML-DE ranked first among all the compared algorithms, such as JADE, RSDE, NBOLDE, EAPSO, MGDE, ACEPSO, FPSO, DEAH, CEO, and CPO, on the CEC2017 test functions for both 30D and 50D problems, and achieved the top rank in FS in the metrics, such as MCE, precision, recall, and F1-score in feature selection problems, confirming its effectiveness across diverse categories of optimization problems.</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

Diversified multi-learning strategy with fitness state-based partitioning in differential evolution algorithm for feature selection

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Deepak Sahu,
  • Shubham Gupta

摘要

The Differential Evolution (DE) algorithm is a well-known evolutionary algorithm to deal with global optimization problems. However, it shows the shortcomings of inefficient diversity and slow convergence while dealing with complex problems. In this paper, we propose a modified version of the DE, denoted as DML-DE, by integrating with several search strategies, including population partitioning, a multi-learning strategy based on a new guiding vector, and statistical distributions and successful historical information-based parameter settings. The population partitioning is applied to categorize the individuals based on their fitness so that appropriate search rules can be assigned to them. The multi-learning strategy defines three different mutation rules to ensure sufficient and balanced levels of diversity and convergence speed. The DML-DE is validated on standard CEC2017 problems to verify its efficiency in solving continuous optimization problems, and later, extended to its binary version using a transfer function to find the highly effective feature subsets for the feature selection (FS) problems. A collection of 24 datasets from the UCI repository is used to analyze its performance. The results obtained by DML-DE are compared with 11 metaheuristics using various evaluation measures, which demonstrate that DML-DE ranked first among all the compared algorithms, such as JADE, RSDE, NBOLDE, EAPSO, MGDE, ACEPSO, FPSO, DEAH, CEO, and CPO, on the CEC2017 test functions for both 30D and 50D problems, and achieved the top rank in FS in the metrics, such as MCE, precision, recall, and F1-score in feature selection problems, confirming its effectiveness across diverse categories of optimization problems.