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 ›› 2025, Vol. 33 ›› Issue (6): 222-232.doi: 10.16381/j.cnki.issn1003-207x.2022.1410cstr: 32146.14/j.cnki.issn1003-207x.2022.1410

• • 上一篇    下一篇

基于需求学习的数字化鲜活品拍卖平台批量优化策略研究

孔祥天瑞1, 龙艳红1, 于凯泽2, 弋泽龙1, 晏鹏宇2()   

  1. 1.深圳大学经济学院,广东 深圳 518000
    2.电子科技大学经济与管理学院,四川 成都 611731
  • 收稿日期:2022-06-28 修回日期:2023-01-29 出版日期:2025-06-25 发布日期:2025-07-04
  • 通讯作者: 晏鹏宇 E-mail:yanpy@uestc.edu.cn
  • 基金资助:
    国家自然科学基金项目(72031004);教育部人文社会科学研究青年基金项目(22YJC630052);国家社会科学基金重大项目(20&ZD084)

Optimal Lot-sizing for Digital Perishable Auction Platform with Demand Learning

Xiangtianrui Kong1, Yanhong Long1, Kaize Yu2, Zelong Yi1, Pengyu Yan2()   

  1. 1.College of Economics,Shenzhen University,Shenzhen 518000,China
    2.School of Management and Economics,University of Electronic Science and Technology of China,Chengdu 611731,China
  • Received:2022-06-28 Revised:2023-01-29 Online:2025-06-25 Published:2025-07-04
  • Contact: Pengyu Yan E-mail:yanpy@uestc.edu.cn

摘要:

由于鲜活品市场具有供需波动大、交易量大和时效性强等特点,导致不同参拍批量与顺序下的拍品成交价格相差较大,拍卖平台收益难以从整体上达到最优。实践中,拍卖平台采用基于拍卖师个人经验的固化参拍计划决策,难以适应需求信息未知且动态变化的复杂情形。本文从拍卖平台视角,探究如何动态调整拍卖批量从而提升拍卖平台的长期总收益。首先,分析竞拍者之间博弈对均衡竞价的影响,得出给定拍卖批量下的竞拍者均衡竞价策略;随后,基于特定竞拍者估价分布和需求分布,建立拍卖平台的单阶段收益函数;最后,基于需求信息逐步揭示学习,建立基于贝叶斯-马尔科夫决策过程的多阶段收益优化模型,应用滚动时域在线优化方法求解该模型。通过大量数值实验,获得的管理启示为:(1)基于需求信息的不断更新学习,通过动态设定合适的拍卖批量,可以有效提升拍卖收益,尤其在供过于求和需求季节性波动的时期。(2)在初始信念与真实分布差距相同时,需求被低估下进行信息学习的收益提升比例高于需求被高估下进行信息学习的收益提升比例。因此,当拍卖平台难以预测需求时,低估需求下的决策要优于高估需求下的决策。(3)拍卖平台可以通过同时动态调节最优拍卖总阶段、最优保留价、最优拍卖批量来提升总收益。

关键词: 多物品序贯拍卖, 拍卖变量决策, 需求学习, 贝叶斯-马尔科夫决策过程

Abstract:

Due to the characteristics of large fluctuations in supply and demand, large transaction volume and strong timeliness in the fresh product market, the transaction prices of the auction items under different auction batches and sequences are quite different, and it is difficult for the auction platform to achieve the optimal revenue. In practice, the auction platform adopts a solidified auction plan decision based on the auctioneer's personal experience, which is difficult to adapt to complex situations where demand information is unknown and dynamically changing. From the perspective of the auction platform, this article explores how to dynamically adjust the auction batch to improve the long-term total revenue of the auction platform. Firstly, the influence of the game between bidders on the equilibrium bidding is analyzed, and the equilibrium bidding strategy of the bidders under the given auction batch is obtained; secondly, based on the price distribution and demand distribution of specific bidders, the single-stage revenue function of the auction platform is established; finally, a multi-stage revenue optimization model based on Bayesian-Markov decision process is built while dynamically learning demand signal. The online rolling horizon optimization method is applied to solve the model. Through a large number of numerical experiments, the management inspirations are obtained as follows: (1) Based on the continuous updating and learning of demand information, the auction revenue can be effectively improved by dynamically setting the appropriate auction batch, especially in the period of oversupply and seasonal fluctuation of demand; (2) When the gap between the initial belief and the true distribution is the same, the benefit of information learning under underestimated demand is higher than that of information learning under overestimated demand. Therefore, when it is difficult to predict the demand, the decision made by auction platform under the underestimated demand is better than the overestimated demand; (3) The auction platform can dynamically configure the optimal total auction stage, the optimal reserve price, and the optimal auction batch simultaneously to increase total revenue of auction platform.

Key words: multi-item sequential auction, auction variable decision, demand learning, Bayesian-Markov decision process

中图分类号: