Scholarlink

<p>This paper proposes a confidence-weighted adaptive hybrid prediction framework for dynamic multi-objective evolutionary algorithms, termed AHPF-DMOEA, to improve population re-initialization under heterogeneous and severe environmental changes in dynamic multi-objective optimization problems (DMOPs). The proposed framework establishes a closed-loop pipeline consisting of four stages: perception–prediction–decision–generation. In the perception stage, an environment change intensity detection (ECID) module quantifies environmental disturbances and outputs a continuous intensity indicator <i>I</i>(<i>t</i>). In the prediction stage, guided by <i>I</i>(<i>t</i>), a complementary dual-branch predictor is employed to capture both global population drift and critical region (knee) transitions. In the decision stage, an adaptive weight control (AWC) mechanism performs confidence-aware fusion by dynamically weighting the two branches based on their confidence and segment-level performance feedback, thereby reducing the risk of performance degradation due to bias from a single prediction branch under heterogeneous dynamics. In the generation stage, an intensity-aware hybrid initial population generation strategy constructs the next population via multi-channel quota allocation, combining model-guided samples, historical information, and random exploration, enabling fast yet stable re-convergence. The proposed algorithm is evaluated on the DF1–DF14 benchmark suite under four dynamic configurations and compared with representative state-of-the-art baselines. Experimental results demonstrate that AHPF-DMOEA achieves consistently competitive or superior performance, indicating improved response speed, solution quality, and robustness in complex dynamic environments. Because dynamic multi-objective optimization requires repeated population re-evaluation and rapid recovery after environmental changes, the proposed framework provides a computational basis for real-time optimization scenarios and scalable parallel or distributed high-performance computing environments.</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

A confidence-weighted adaptive hybrid prediction framework for dynamic multi-objective evolutionary optimization

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Yue He,
  • Yong Chen,
  • Kun Liang,
  • Yuning Guo

摘要

This paper proposes a confidence-weighted adaptive hybrid prediction framework for dynamic multi-objective evolutionary algorithms, termed AHPF-DMOEA, to improve population re-initialization under heterogeneous and severe environmental changes in dynamic multi-objective optimization problems (DMOPs). The proposed framework establishes a closed-loop pipeline consisting of four stages: perception–prediction–decision–generation. In the perception stage, an environment change intensity detection (ECID) module quantifies environmental disturbances and outputs a continuous intensity indicator I(t). In the prediction stage, guided by I(t), a complementary dual-branch predictor is employed to capture both global population drift and critical region (knee) transitions. In the decision stage, an adaptive weight control (AWC) mechanism performs confidence-aware fusion by dynamically weighting the two branches based on their confidence and segment-level performance feedback, thereby reducing the risk of performance degradation due to bias from a single prediction branch under heterogeneous dynamics. In the generation stage, an intensity-aware hybrid initial population generation strategy constructs the next population via multi-channel quota allocation, combining model-guided samples, historical information, and random exploration, enabling fast yet stable re-convergence. The proposed algorithm is evaluated on the DF1–DF14 benchmark suite under four dynamic configurations and compared with representative state-of-the-art baselines. Experimental results demonstrate that AHPF-DMOEA achieves consistently competitive or superior performance, indicating improved response speed, solution quality, and robustness in complex dynamic environments. Because dynamic multi-objective optimization requires repeated population re-evaluation and rapid recovery after environmental changes, the proposed framework provides a computational basis for real-time optimization scenarios and scalable parallel or distributed high-performance computing environments.