发现信息

  • 作者:Pengfei Zhou, Junli Liang, Jiayuan Chen, Rong Lin, Ziwen Wang, Zhang, Ping, Wangqiu Zhou, Qi Song
  • 年份:2026
  • 发表源:Zenodo (CERN European Organization for Nuclear Research)
  • OpenAlex ID:https://openalex.org/W7163890437
  • DOI:https://doi.org/10.5281/zenodo.20594596
  • arXiv:待核验
  • 检索主题:self-evolving
  • OpenAlex 引用数快照:0(2026-11-12)

摘要线索

In time-series forecasting, architectures tailored to domain-specific dataset often outperform generic models. However, manually designing such architectures requires substantial domain knowledge and engineering effort. To address this challenge, this paper proposes the Self-Evolving Time-Series Agent (SE-TSA), a large language model-driven programmatic evolution framework that iteratively creates and refines forecasting architectures without explicit human intervention in architecture design. SE-TSA incorporates a dual-strategy generation mechanism (EXPLORE and EXPLOIT), adopts the MAP-Elites quality-diversity framework to encourage architectural diversity and mitigate premature convergence, and employs a reflection mechanism to accumulate interpretable design principles. We conduct extensive experiments on nine cross-domain datasets, including domain-transfer analysis and ablation studies, demonstrating the effectiveness of the proposed method compared to diverse kinds of baselines. The source code is available at https://github.com/mumiao2000/SE-TSA/.

核验与筛选

  • 当前状态:待筛选、去重和一手来源核验。
  • OpenAlex 元数据和引用数只用于发现,不作为正式引文或质量判断。
  • 记录正式版本、venue、方法细节、实验协议和局限后,才迁入稳定论文卡片。