Scholarlink

<p>Over the past two decades, agile development has become a mainstream software engineering paradigm due to its flexibility and iterative nature. However, accurate cost estimation in agile projects remains challenging, mainly due to frequent requirement changes and data scarcity. Traditional estimation methods and existing machine learning models often fail to adapt effectively to the dynamic agile environment. To address these issues, this paper proposes a novel cost estimation model combining Kolmogorov-Arnold Networks (KAN) with the AdamW optimizer. KAN captures complex nonlinear relationships through hierarchical polynomial mapping, making it suitable for modeling dynamic cost variations. AdamW improves convergence speed and stability with adaptive learning rates and momentum mechanisms. To alleviate data scarcity, the SMOTE-NC technique is applied to generate 1000 synthetic samples based on 75 actual agile project data. K-fold cross-validation is used to enhance the model’s generalization ability. Experimental results demonstrate that the proposed KAN-AdamW model achieves superior performance, with a Mean Absolute Error (MAE) of 11,504.08 and a Mean Relative Error (MRE) of 0.12-outperforming traditional Artificial Neural Networks (ANN) and function point-based models. The model also shows strong performance in accuracy and R-squared metrics, indicating high predictive stability and precision. This study offers a data-driven and effective solution for agile cost estimation and provides empirical support for addressing data limitations using SMOTE-NC. Furthermore, it highlights the potential of KAN for broader applications in cost modeling.</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 data-driven cost estimation model for agile development based on Kolmogorov-Arnold Networks and AdamW optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月23日

查看全文
  • Xiaoyan Zhao,
  • Xin Xiong,
  • Zulkefli Mansor,
  • Rozilawati Razali,
  • Mohd Zakree Ahmad Nazri,
  • Liangyu Li

摘要

Over the past two decades, agile development has become a mainstream software engineering paradigm due to its flexibility and iterative nature. However, accurate cost estimation in agile projects remains challenging, mainly due to frequent requirement changes and data scarcity. Traditional estimation methods and existing machine learning models often fail to adapt effectively to the dynamic agile environment. To address these issues, this paper proposes a novel cost estimation model combining Kolmogorov-Arnold Networks (KAN) with the AdamW optimizer. KAN captures complex nonlinear relationships through hierarchical polynomial mapping, making it suitable for modeling dynamic cost variations. AdamW improves convergence speed and stability with adaptive learning rates and momentum mechanisms. To alleviate data scarcity, the SMOTE-NC technique is applied to generate 1000 synthetic samples based on 75 actual agile project data. K-fold cross-validation is used to enhance the model’s generalization ability. Experimental results demonstrate that the proposed KAN-AdamW model achieves superior performance, with a Mean Absolute Error (MAE) of 11,504.08 and a Mean Relative Error (MRE) of 0.12-outperforming traditional Artificial Neural Networks (ANN) and function point-based models. The model also shows strong performance in accuracy and R-squared metrics, indicating high predictive stability and precision. This study offers a data-driven and effective solution for agile cost estimation and provides empirical support for addressing data limitations using SMOTE-NC. Furthermore, it highlights the potential of KAN for broader applications in cost modeling.