美国对外战争计量研究——美国战争史量化研究及战争行为预测

Abstract

战争，是人类的群体暴力行为，是实现政治和外交目的的极端手段。这种极端手段会在什么时间、什么地点、什么条件下使用一直是各国家、集团和组织的决策者们希望掌握的重要情报，但提前获得这样的情报难度较大。从公开发表来看，国内外大部分专家、学者对战争的预判大多相对主观，可见要在纷繁复杂的因素中探寻战争决策的有效规律对于任何研究者来说都是挑战。那么，战争的规律是否真的存在？笔者通过对历史文献的回顾发现，人类的群体暴力行为是物质世界长期发展的结果，而黩武主义（militarism）在美国国家性格中长期存在并且持续表达，这种客观性和长期性意味着美国的战争行为一定具有可以被认知的理性成分，问题在于怎样找到并把握它。

本文创新性地通过观测美国“战争机器”的各种“信号”来预测战争。为了实现这一研究目标，笔者基于系统工程思维，借助前沿数据科学方法，深入研究了美国发动或参与战争的行为能否被定量预测以及怎样预测，完成了前后递进的5项任务：一是美国战争史分期。运用大数据方法研究美国自建国以来的军事活动历史，绘制出美国军事活跃度曲线（U.S. Military Activeness Curve, USMAC），并根据USMAC的峰值、谷值和背景军事活跃程度（Background Military Activeness, BMA）将美国军事活动划分为相互独立的5个阶段，为后续研究建立起对美国战争的历史性、阶段性认识。二是美国战争分类。在前人研究基础上，运用机器学习技术建立对美国历史上若干军事行动的分类模型，将战争门槛从单一、主观、固定的经验值，转化为多维、客观、动态的算法，并用模型对美国历史上若干军事行动进行了分类，从中筛选出16个典型战争案例，为后续研究确定了一个大小适中、代表性强的样本。三是美国战争典型案例分析。将16次典型战争放入5个阶段，着重运用定性研究方法考察历次美国战争的背景、原因和战争决策过程，提炼出与美国发动或参与战争过程中的若干共性因素，并对这些因素进行讨论，为后续定量研究打下定性研究基础。四是美国战争相关因素数据实验。运用数理统计、自然语言处理等技术对美国战争或冲突若干相关变量的解释性进行实验并获得显著结果，一定程度上证明了对战争进行量化预测的可行性，为后续数据建模工程提供重要指引。五是美国战争预测。基于前述研究成果，构建美国战争决策空间（USWDMIS），搭建人工神经网络，经过百万次迭代，训练出能够稳定预测美国对外武装冲突行为的数学模型，并对1970年以来的实际冲突案例进行了准确度较高的回测

研究从理论角度证明美国发动或参与战争的行为在一定程度上能够被预测，并从实践角度实现了预测。

Alternate abstract:

War, as a collective act of violence by humanity, represents an extreme measure employed to achieve political and diplomatic aims. The utilization of such extreme measures, in terms of timing, location, and conditions, has always been crucial information sought after by decision-makers in nations, groups, and organizations. However, obtaining this information in advance presents significant challenges. From the perspectives of experts and scholars, the prediction of wars, based on publicly available sources, tends to be largely subjective. It is evident that uncovering effective patterns governing the decision-making behind war amidst the multitude of intricate factors presents a challenge for any researcher. Therefore, does a discernible pattern truly exist when it comes to war? Through a retrospective analysis of historical literature, the author discovered that collective acts of violence by humans are the long-term outcome of world development. Additionally, militarism has persistently and objectively existed as a component of the American national character. This objectivity and long-standing nature imply that American warfare necessarily entails recognizable rational elements. The question then becomes how to identify and harness those elements.

In an innovative approach, this article predicts wars through the observation of various "signals" emitted by the American "war machine." To achieve this research goal, the author applies Systems Engineering thinking and leverages cutting-edge data science methods to delve deeply into the quantification and prediction of American initiation or involvement in wars. This progressive research consists of FIVE major tasks. Firstly, it involves the delineation of stages in American war history. Utilizing Big Data approaches, the author investigates the historical military activities of the United States since its founding, generating the U.S. Military Activeness Curve (USMAC). Based on the peaks, troughs, and background military activity levels (Background Military Activeness, BMA) of the USMAC, the author divides American military activities into five distinct stages. This provides a historical and phased understanding of American warfare to establish a base for subsequent research. Secondly, it comprises the categorization of American wars. Building upon prior research, the author employs Machine Learning techniques to establish a classification model for several military actions in American history. This model transforms the threshold for war from single, subjective, and fixed empirical viewpoints into a multidimensional, objective, and dynamic algorithm. By applying this model to classify various military actions in American history, the author identifies 16 representative war cases, constituting a moderately sized and highly representative sample set for further study. Thirdly, the research conducts an analysis of typical American war cases. Placing the 16 representative wars within the five stages, the author extensively employs qualitative research methods to examine the background, causes, and decision-making processes behind each American war. Common factors prevalent in American initiation or involvement in wars are distilled and discussed, laying a qualitative research foundation for subsequent quantitative studies. Fourthly, it involves data experiments concerning factors related to American wars. Leveraging techniques such as mathematical statistics and natural language processing, the author conducts experiments to examine the explanatory power of several variables concerning American wars or conflicts. These experiments yield significant results, substantiating the feasibility of quantitatively predicting wars and providing crucial guidance for subsequent data modeling efforts. Lastly, the research focuses on war prediction in the American context. Building upon the aforementioned research achievements, the author constructs the U.S. War Decision-Making Information Space (USWDMIS) and establishes an artificial neural network. Through millions of iterations, a stable mathematical model capable of predicting American armed conflicts is trained, exhibiting a high level of accuracy when retroactively tested against conflict cases since 1970.

The research both theoretically demonstrates the predictability, to a certain extent, of American initiation or involvement in wars and practically realizes such predictions.