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 ›› 2026, Vol. 34 ›› Issue (3): 357-368.doi: 10.16381/j.cnki.issn1003-207x.2024.2237cstr: 32146.14.j.cnki.issn1003-207x.2024.2237

• • 上一篇    

“结构-信息”耦合网络下基于强化学习的群智共识决策建模研究

谭笑, 蔡嘉麒, 巩在武()   

  1. 南京信息工程大学管理工程学院,江苏 南京 210044
  • 收稿日期:2024-12-12 修回日期:2025-05-25 出版日期:2026-03-25 发布日期:2026-03-06
  • 通讯作者: 巩在武 E-mail:zwgong26@163.com
  • 基金资助:
    国家自然科学基金项目(72501146);国家自然科学基金项目(72371137);教育部人文社会科学研究青年基金项目(24YJC630191);江苏省社会科学基金青年项目(24JLC030);中国博士后科学基金项目(2024M751483);江苏省高校哲学社会科学一般项目(2023SJYB0178)

Group Intelligence Consensus Decision Modeling Based on Reinforcement Learning with theStructure-InformationCoupled Network

Xiao Tan, Jiaqi Cai, Zaiwu Gong()   

  1. School of Management Science and Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China
  • Received:2024-12-12 Revised:2025-05-25 Online:2026-03-25 Published:2026-03-06
  • Contact: Zaiwu Gong E-mail:zwgong26@163.com

摘要:

人工智能与信息技术的快速发展,推动协同群决策逐渐成为解决应急管理、工程论证等复杂系统决策问题的关键手段,其规模化、社交化、动态化特征对决策模型自适应能力提出了根本性需求。然而,非共识逐次辨识的传统范式难以满足群偏好的高效聚合与动态环境实时响应,亟需探索自适应的智能化群决策方法。考虑到决策成员的信任结构与偏好信息能够揭示个体间的潜在关联和共识基础,二者在动态交互中的耦合效应则直接影响群决策过程,因此,本文提出一种“结构-信息”耦合网络下基于强化学习的群智共识决策模型。首先,量化决策成员间有向信任关系与偏好信息相似性,构建结构网络与信息网络,并分析其耦合交互机制。进一步,设计融合群体共识水平、协同激励与协同成本的综合奖励函数,针对耦合网络下群决策问题中决策行为(动作空间)的离散特性以及决策环境(状态空间)的连续特性,引入深度Q网络(deep Q-network, DQN)算法实现自适应的共识策略学习。最后,仿真实验表明所提模型在提高共识达成效率的同时,能够有效降低协同成本,并通过参数敏感性分析验证了模型在差异化行为机制中的适用性。

关键词: 群智协同, 共识决策, DQN算法, 社会网络分析, 偏好演化

Abstract:

The rapid development of artificial intelligence and information technology has propelled collaborative group decision-making to become a key approach for addressing complex systemic decision-making problems in emergency management, engineering demonstration, and other fields. Its characteristics of scalability, social interactivity, and dynamism fundamentally necessitate adaptive capabilities in decision-making models. However, the traditional paradigm of non-consensus sequential identification struggles to meet the demands of efficient group preference aggregation and real-time response in dynamic environments, urgently necessitating the exploration of adaptive intelligent group decision-making methods. Considering that the trust structure and preference information of decision-makers can reveal the latent inter-individual relationships and consensus foundation, and their coupled effects during dynamic interactions are recognized as directly influencing the group decision-making process. Therefore, a reinforcement learning-based intelligent group consensus decision-making model with the “structure-information” coupled network is proposed in this study. First, trust relationships and preference information similarity among decision-makers are quantified to construct the structure network and the information network, and the coupling interaction mechanisms of these two networks are then investigated. Furthermore, an integrated reward function with group consensus level, collaborative incentives, and collaborative costs is designed. To address the discrete characteristics of decision-making behaviors (action space) and the continuous characteristics of decision environments (state space) in group decision-making with the coupled network, the Deep Q-Network (DQN) algorithm is introduced to adaptively optimize consensus strategies. Finally, simulation experiments demonstrate that the proposed model effectively enhances consensus-reaching efficiency while reducing collaborative costs, and its applicability in differentiated behavioral mechanisms is verified through parameter sensitivity analysis.

Key words: collaborative intelligence, consensus decision-making, DQN algorithm, social network analysis, preference evolution

中图分类号: