基于多智能体强化学习的施工班组协同任务分配方法OA
Collaborative Task Allocation Method for Construction Crews Based on Multi-agent Reinforcement Learning
针对施工过程中班组任务分配能力差、协调性不足的问题,提出了基于多智能体近端策略优化(MAPPO)的施工班组协同任务分配方法.首先,采用Unity 3D仿真引擎搭建多智能体施工过程仿真环境;其次,建立任务分配马尔科夫决策过程,并设计了基于MAPPO的多班组协同决策和分布式执行框架;最后,提出了施工任务有效动作筛选(VAF)机制和基于进度势函数的奖励设计(PBRS)方法,实现了智能体对环境的高效探索,解决了任务分配的可变动作空间和稀疏奖励问题.案例仿真结果表明,本文方法相比于传统多智能体方法工期缩短14.8%,工时利用率提升9.7%.进一步地,将MAPPO与独立近端策略优化(IPPO)进行对比,验证了协同决策对于任务分配的必要性.
To address the poor allocation capability and weak coordination among crews during construction,a multi-agent proximal policy optimization(MAPPO)-based method for collaborative task allocation was proposed.A multi-agent construction simulation environment was first developed using Unity 3D.A Markov decision process model was then formulated,and a MAPPO-based framework with centralized training and decentralized execution was designed.Additionally,a valid action filtering(VAF)mechanism for construction tasks and a progress-based reward shaping(PBRS)method were proposed,enabling the efficient environmental exploration by agents and addressing the challenges of variable action spaces and sparse rewards in task allocation.Case study results demonstrate that the proposed method reduces the project duration by 14.8%and improves the labor utilization by 9.7%compared to conventional multi-agent methods.Furthermore,comparative analysis between MAPPO and independent proximal policy optimization(IPPO)demonstrates the necessity of collaborative decision-making for task allocation.
王一凡;杨彬;汪丛军;刘伯达
同济大学 土木工程学院,上海 200092同济大学 土木工程学院,上海 200092中亿丰建设集团股份有限公司,江苏 苏州 215131同济大学 土木工程学院,上海 200092
建筑与水利
土木工程施工施工仿真任务分配多智能体强化学习多智能体系统
civil engineering constructionconstruction simulationtask allocationmulti-agent reinforcement learningmulti-agent system
《同济大学学报（自然科学版）》 2026 (7)
1005-1014,10
国家重点研发计划(2024YFD1600402)
评论