支持重规划的战时保障动态调度研究  

Research of Dynamic Scheduling With Re-planning for Wartime Logistics Support

作　　者：曾斌[1] 樊旭 李厚朴[2] ZENG Bin;FAN Xu;LI Hou-Pu(Department of Management and Economics,Naval University of Engineering,Wuhan 430033;College of Electrical Engineering,Naval University of Engineering,Wuhan 430033)

机构地区：[1]海军工程大学管理工程与装备经济系,武汉430033 [2]海军工程大学电气工程学院,武汉430033

出　　处：《自动化学报》2023年第7期1519-1529,共11页Acta Automatica Sinica

基　　金：国家优秀青年科学基金(42122025);湖北省杰出青年科学基金(2019CFA086)资助。

摘　　要：复杂多变的战场环境要求后装保障能够根据战场环境变化,预见性地做出决策.为此,提出基于强化学习的动态调度方法.为准确描述保障调度问题,提出支持抢占调度、重分配及重部署决策的马尔科夫决策过程(Markov decision process,MDP)模型,模型中综合考量了任务排队、保障优先级以及油料约束等诸多问题的影响;随后设计改进策略迭代算法,训练基于神经网络的保障调度模型;训练后的神经网络模型能够近似计算状态价值函数,从而求解出产生最大期望价值的优化调度策略.最后设计一个分布式战场保障仿真实验,通过与常规调度策略的对比,验证了动态调度算法具有良好的自适应性和自主学习能力,能够根据历史数据和当前态势预判后续变化,并重新规划和配置保障资源的调度方案.It is necessary for the logistics and equipment support to make decision according to battlefield environment changes in the complicated and changeable battlefield.A dynamic scheduling method is proposed to overcome the problem.Firstly,a Markov decision process(MDP)model is proposed to formulate the preemptive scheduling,reassignment and redeployment of wartime logistics support force,which considers the impact of task queuing,maintenance priorities and oil limits,etc.Secondly,an improved policy iterative algorithm is designed to train the scheduling model based on neural network.The trained neural network model can approximate the state value function and solve the optimal scheduling decision resulting in the best expected value.Finally,a distributed logistics support experiment is designed to verify the applicability of dynamic scheduling algorithm compared to the traditional scheduling scheme,the results show that the algorithm has the capabilities of adjusting and re-planning the scheduling scheme according to historical data and prediction based on current situation.

关 键 词：战时保障 重规划 马尔科夫决策过程 动态调度 强化学习 

分 类 号：E91[军事] TP18[自动化与计算机技术—控制理论与控制工程]