【作者】 石崇林；

【导师】 张茂军；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2012， 博士

【摘要】 计算机兵棋系统具备节省训练经费、不受场地限制、接近真实体验等优点，是和平时期战略战役指挥能力训练的有效途径，其发展也越来越受到国内外的重视。对推演数据进行分析是计算机兵棋系统中的一个重要组成部分，可以帮助指挥员找出推演过程中发挥的长处和暴露的不足，充分发挥计算机兵棋系统在指挥能力培养和作战理论探索上的优势；而同时，计算机兵棋系统的大规模、多尺度、高复杂性等特点决定了对推演数据进行分析又是一个难点问题。本文正是在这一背景下，以大型计算机兵棋系统的军事需求为牵引，提出了基于数据挖掘的数据分析方法，论文研究成果可归纳如下：（1）考虑到兵棋系统结构的复杂性和高性能的运行要求，以及兵棋数据的多样性和大规模性，在分析兵棋系统数据特点的基础上，描述了三种兵棋数据结构，提出了兵棋数据的采集原则，定义了七种基本采集模式，然后根据这些采集模式系统分析和比较了当前主要的采集方法和技术的优缺点及其适用范围，满足了兵棋系统在不同开发阶段的采集需求；提出了三种基于文件的兵棋推演数据存储策略，并依据时空复杂度计算方法对存储策略的访问和存储效率进行了对比分析，较好地解决了高效的数据存储需求等开发与应用中的一些实际问题。（2）为了能够快速有效地检测兵棋推演过程中对抗各方在兵力部署上的缺陷，以节省指挥员分析战场态势的时间，使其有更多的精力花在施计用谋和排兵布阵上，提出了一种基于密度的QDBSCAN算法。QDBSCAN算法对传统的DBSCAN算法做了一些改进：①在邻近度度量上提出了最短可行路径的概念，使聚类更符合计算机兵棋的规则；②动态设置密度参数，不仅参数的设置方式让指挥员容易接受，同时也使算法的结果更能被指挥员所理解；③采用提出的代表对象选择方法来减少对对象邻域的判断次数，并按区域对数据进行分组，在缩小聚类规模的同时也减少了最短可行路径的计算时间，进一步提高了算法的效率。实验表明，QDBSCAN算法的性能在数据规模较大的情况下，明显优于DBSCAN算法。（3）通过热区定位，可以帮助指挥员整体审视战场态势，辅助分析对抗方的作战意图。针对计算机兵棋地图和推演数据的特点，结合图论的思想，提出了一种基于连通树的热区检测算法，可检测任意形状的热区。首先从数学的角度明确了热区相关的几个定义，并证明了热区存在于每个最小化分的连通区域中。在此基础上，提出了先构建连通树将数据集按连通区域进行最小划分，再对每棵连通子树进行剪枝处理。通过调整密度阈值可以改变指挥员的观察视角，以适合不同层次指挥员的需要。理论和实验结果均表明了算法的有效性。（4）通过轨迹聚类可以发现兵棋推演作战实体之间的整体行为模式，进而可以辅助判断指挥员的作战意图。因此，在充分分析作战实体轨迹数据特点的基础上，提出了计算机兵棋作战实体轨迹聚类CTECW算法。整个算法由三个部分组成：轨迹预处理，轨迹分段聚类以及可视化表现。在轨迹预处理阶段，先将实体原始轨迹转化成实体简化轨迹，再进一步处理成轨迹分段；在轨迹分段聚类阶段，引入了DENCLUE算法中密度函数的概念，并基于DBSCAN算法的基本框架采用提出的相似性度量函数对轨迹分段进行聚类，并对参数的优选设置进行了讨论；在可视化表现阶段，将轨迹分段聚类的结果以赋有军事涵义的形式展现给指挥员，使得指挥员对聚类的结果更容易理解和接受。理论和实验结果均表明了算法的有效性。（5）在分析了数据挖掘系统的发展以及主流数据挖掘系统的体系结构和功能的基础上，设计并实现了基于数据挖掘技术的兵棋推演数据分析原型系统。更多还原

【Abstract】 In peacetime, the computer wargame is one of the most suitable tools for higherlevels of military command training due to its superiority in terms of money, amount ofsubordinate troops required, land use and approximative realistic experience. And itsdevelopment is receiving increased emphasis. As an important component of thecomputer wargame, analyzing wargaming data can help commanders to findstrongpoints or weakness exploited in the wargaming process, improving its effect inmilitary command training and operation theories exploration. At the same time,analyzing wargaming data is also a difficult problem because of that the computerwargame possesses such characteristics as large-scope, multi-scale, high complexity,and so on. It is under such background that the author effectively studies thecorresponding martial requirements of the computer wargame and proposes some dataanalysis methods based on data mining, the achievements of this dissertation can beconcluded as follows:(1) Considing the complex structure and high running performance of the computerwargame as well as the diversity and massiveness of wargaming data, based on studyingthe characters of the wargaming data, the author describes three kinds of wargamingdata structure, proposes data collection principles, defines seven basic collection modes,and then analyses forth and compares the merits and defects of current main datacollection methods and technologies and their applicability according to those definedcollection modes, which meets data collection requiements at various developmentstages. Besides that, to solve some practical problems such as efficient data storageneeds in the development and applications of the computer wargame, we provide threedata storage policies based on files and their efficiencies of access and storage arecomparatively analysed using the spatial-temporal complexity computing method.(2) A clustering algorithm named QDBSCAN is proposed for determining thevulnerability of units’ deployment rapidly by detecting the isolated units in order to savethe time consumed by commaders for analysing the battlefield situation in thewargaming process. Compared with DBSCAN, QDBSCAN made some improvementsin such aspects:①defined the shortest viable path as the similarity measurement tomake the clustering algorithm more coincident with the rules of computer wargames;②set the density parameters dynamically instead of statically;③chose a small numberof representative objects to expand the cluster so that the execution frequency of regionquery was reduced and grouped the whole dataset by divisiory regions in order toreduce the scale of clustering and ulteriorly enhance the efficiency of the algorithm.Experimental results indicate that QDBSCAN is more effective and efficient thanDBSCAN in clustering large datasets. (3) A hotspot detection algorithm based on connected tree is proposed, which iscapable of detecting arbitrarily shaped hotspots during the wargaming process. Bydetecting the areas with high concentrations of martial events, this algorithm couldassist commaders understanding the whole wargaming battlefield situation andconjecturing the intent of opponents. After making the definition of a hotspot, aconnected tree is built in order to least divide the whole dataset into connected regions,and a pruning procedure is carried out according to the provided density threshold value.Each pruned connected subtree is a hotspot which we would like to acquire. Throughadjusting the density threshold, commanders with different levels can chooseappropriate observational views. Both the theoretical analysis and experimental resultsverify the effectiveness of the algorithm.(4) By clustering trajectories of the wargame entities, we may find their holisticbehavioral patterns which could help commanders estimate the intent of their opponents.Thereby, in allusion to the characters of trajectory data of the wargame entities atrajectory clustering algorithm named CTECW is proposed. This algorithm is composedof three parts: trajectory pretreatment, trajectory segments clustering and visualpresentation. Trajectory pretreatment transforms original trajectories into simplifiedones which are ulteriorly processed into linear segments. In the second part, the conceptof density function derived from DENCLUE is introduced and trajectory segments areclustered based on our own similarity measure under the framework of DBSCAN.Moreover, how to choose optimal parameters is discussed detailedly. Visualpresentation exhibits clustering results with martial meanings which can be easilycomprehended by commanders. Both the theoretical analysis and experimental resultsverify the effectiveness of the algorithm.(5) After analysing the development of the data mining system as well asarchitectures and functions of main data mining systems, a wargame data analysisprototype system based on data mining is designed.更多还原