主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2024, Vol. 32 ›› Issue (1): 260-267.doi: 10.16381/j.cnki.issn1003-207x.2020.0816

• • 上一篇    

基于XGBoost算法的数据驱动单周期报童问题研究

严雨婷,毕文杰()   

  1. 中南大学商学院,湖南 长沙,410083
  • 收稿日期:2020-05-08 修回日期:2020-10-20 出版日期:2024-01-25 发布日期:2024-02-08
  • 通讯作者: 毕文杰 E-mail:beenjoy@126.com
  • 基金资助:
    国家社会科学基金项目(23BJL126)

A Data-driven Single-Period Newsvendor Problem Based on XGBoost Algorithm

Yuting Yan,Wenjie Bi()   

  1. School of Business,Central South University,Changsha 410083,China
  • Received:2020-05-08 Revised:2020-10-20 Online:2024-01-25 Published:2024-02-08
  • Contact: Wenjie Bi E-mail:beenjoy@126.com

摘要:

求解单周期报童问题通常假设需求分布服从某一特定形式,通过历史数据预测需求并求解优化模型。尽管此假设可以简化分析,但它并不能反映随时间动态变化的真实需求分布,且不可避免地将历史估计预测误差传递给优化过程。为解决这一问题,本文提出了一种将估计和优化相结合的方法。该方法采用数据驱动计算权重的形式来处理与特征相关的需求不确定性,并引入XGBoost算法对具有非平稳需求的单周期报童问题进行求解。在观察到影响需求的特征后,该方法能直接确定最佳库存决策。本文对运营管理领域的贡献主要包括两个方面:1)提供了一种解决方案,将名为XGBoost的可扩展端到端树提升系统的机器学习算法集成到了单周期报童问题中。2)将该集成估计优化算法应用于不同目标服务水平和训练样本量的真实数据集,并与现有研究报童问题的几种标准方法进行比较,结果表明,该方法至少可以降低约5%的库存成本。

关键词: 报童问题, 机器学习, SAA, 分位数回归, 优化

Abstract:

Solving single-period newsvendor problem usually assumes that the demand distribution follows a particular form, predicts demand from historical data, then solves optimization models. Although this assumption simplifies the analysis, it does not reflect the true distribution of demand over time and inevitably passes on historical estimated prediction errors to the optimization process. To solve this problem, a method that combines estimation and optimization is proposed. This method adopts the form of data-driven weights to deal with the demand uncertainty related to features and introduces XGBoost algorithm to solve single- period newsvendor problem with non-stationary demand. After observing the features that affect demand, the method directly determines the optimal inventory decision. This paper's contribution to the field of operations management mainly includes twofold: 1) It integrates a scalable end-to-end tree boosting system called XGBoost into a single-period newsvendor problem. 2) The integrated estimation optimization algorithm is applied to a real-world data set under different target service levels and training sample sizes, and is compared with several standard methods for studying newsvendor problems. The results show that this method can reduce inventory costs by at least about 5%.

Key words: newsvendor, machine learning, sample average approximation, quantile regression, optimization

中图分类号: