AD1181578 Merged Public

February 2022 February 2022
Leveraging AI to Mitigate Civilian Harm 利用人工智能减轻平民伤害

Larry Lewis and Andrew Ilachinski 拉里*刘易斯和安德鲁Ilachinski
DISTRIBUTION STATEMENT A. Approved for public release: distribution unlimited. 发行声明A.批准公开发行:发行无限.

Cleared for public release 获准公开释放
DOP-2021-U-030953-1Rev-2 DOP-2021-U-030953-1Rev-2
Abstract Abstract
Over the last decade, there has been great debate about the introduction of AI to warfare. However, that debate 在过去十年中，关于将人工智能引入战争的争论很大。然而，这场争论主要是关于如何确保人工智能应用在战争
has been primarily about how to make sure that AI applications are not indiscriminate in warfare. There is 中不会不分青红皂白。根据国际法和人道原则，还有另一个重要问题：我们如何利用人工智能保护平民免受伤害
another important question, in light of international law and the principle of humanity: how can we use AI to ？我们如何通过使用这一新兴技术来减少整体上的痛苦、伤害和破坏？这份报告是实现这一目标的具体第一步。
我们发现人工智能可以用来帮助解决伤害模式，从而减少伤害的可能性。然后，我们讨论军队可以优先考虑的一
protect civilians from harm? And how can we lessen the infliction of suffering, injury, and destruction overall
些潜在重点领域，以减少对平民的总体风险。
through the use of this emerging technology? This report represents a concrete first step toward meeting this
goal. We find that AI can be used to help address patterns of harm and thus reduce the likelihood of harm. We
then discuss some potential areas of focus militaries could prioritize in order to reduce risks to civilians overall.
This document contains the best opinion of CNA at the time of issue. 本文件包含了CNA在发布时的最佳意见。
It does not necessarily represent the opinion of the sponsor 它不一定代表发起人的意见
Distribution Distribution
DISTRIBUTION STATEMENT A. Approved for public release: distribution unlimited. 发行声明A.批准公开发行:发行无限.
Public Release. 公开发布。

2/3/2022 2/3/2022
This work was performed under Federal Government Contract No. N00014-16-D-5003. 这项工作是根据联邦政府第1号合同进行的。 N00014-16-D-5003。
Cover image credit: Photo By: Marine Corps Staff Sgt. Jacob Osborne 封面图片来源：照片作者：海军陆战队员雅各布奥斯本中士
Approved by: February 2022 Approved by: February 2022
Dr. David Broyles, Director 戴维*布洛伊博士，董事

Special Activities and Intelligence Program 特别活动及情报计划
Operational Warfighting Division 作战部
Request additional copies of this document through inquiries@cna.org. 透过以下途径索取本文件的额外副本inquiries@cna.org...
Copyright © CNA 2022 版权附图CNA2022

Form Approved 批准表格
REPORT DOCUMENTATION PAGE OMB No. 0704-0188 报告文档页面 OMB No. 0704-0188
Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the 这种信息收集的公开报告负担估计平均每个答复1小时包括审查指示、搜索现有数据源、收集和维护所需数据以及完成和审查这种信息收集的时间。发送有关此负担估计或此信息收集的任何其他方面的意见，包
data needed, and completing and reviewing this collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing 括减少此负担的建议国防部，华盛顿总部服务，信息操作和报告局（0704-0188），1215Jefferson Davis Highway，Suite1204，Arlington，VA222024302。受访者应注意，尽管有任何其他法律规定，如果信息
this burden to Department of Defense, Washington Headquarters Services, Directorate for Information Operations and Reports (0704-0188), 1215 Jefferson Davis Highway, Suite 1204, Arlington, VA 22202- 收集没有显示当前有效的OMB控制号码，任何人都不得因未能遵守信息收集而受到任何处罚。请勿将表格交回上述地址。 1. 报告日期(DD-MM-YYYY)2022年2月
4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to any penalty for failing to comply with a collection of information if it does not display a currently
valid OMB control number. PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS.
1. REPORT DATE (DD-MM-YYYY) 2. REPORT TYPE 3. DATES COVERED (From - To) 2. 报告类型1Rev 3. 涵盖日期(由至)
February 2022 1Rev
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER 4. 标题和副标题(U)利用人工智能减轻平民伤害 5A.合约编号N00014-16-D
(U) Leveraging AI to Mitigate Civilian Harm N00014-16-D-5003 -5003
5b. GRANT NUMBER 5B.补助金编号
5c. PROGRAM ELEMENT NUMBER 5C.程序元素编号0605154N

0605154N
6. AUTHOR(S) 5d. PROJECT NUMBER 6. 作者(小号)拉里*刘易斯,安德鲁Ilachinski 5D.项目编号R0148
Larry Lewis, Andrew Ilachinski R0148
5e. TASK NUMBER 5E.任务编号E239.00
E239.00
5f. WORK UNIT NUMBER 5F.工作单位编号
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION REPORT 7. 执行机构名称及地址 8. 执行组织报告编号
NUMBER
Center for Naval Analyses 海军分析中心3003华盛顿阿灵顿大
3003 Washington Blvd 道，VA22201
Arlington, VA 22201 DOP-2021-U-030953-1Rev DOP-2021-U-030953-1Rev
9. SPONSORING / MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S) 9. 赞助/监察机构名称及地址 10. SPONSOR/MONITOR’S ACRONYM(S)
Office of the Chief of Naval Operations (N803) 海军作战处处长办公室(N803)
2000 Navy Department 2000海军部
Washington, DC 20350 11. SPONSOR/MONITOR’S REPORT Washington, DC 20350 11. SPONSOR/MONITOR’S REPORT
NUMBER(S) NUMBER(S)
12. DISTRIBUTION / AVAILABILITY STATEMENT 12. 发行/可用性声明发行声明A.批准公开发行：发行无限。

DISTRIBUTION STATEMENT A. Approved for public release: distribution unlimited.
13. SUPPLEMENTARY NOTES 13. 补充说明
14. ABSTRACT 14. 摘要在过去十年中，关于将人工智能引入战争的争论非常激烈。

Over the last decade, there has been great debate about the introduction of AI to warfare.
However, that debate has been primarily about how to make sure that AI applications are not 然而，这场争论主要是关于如何确保人工智能应用在战争中不会不分青红皂白。根据国际法和人道原则，还有另一个重要问题：
indiscriminate in warfare. There is another important question, in light of international law 我们如何利用人工智能保护平民免受伤害？我们怎样才能通过使用这一新兴技术来减少整体上的痛苦、伤害和破坏呢？这份报告
and the principle of humanity: how can we use AI to protect civilians from harm? And how can 是实现这一目标的具体第一步。我们发现人工智能可以用来帮助解决伤害模式，从而减少伤害的可能性。然后，我们讨论军队可
we lessen the infliction of suffering, injury, and destruction overall through the use of 以优先考虑的一些潜在重点领域，以减少对平民的总体风险。
this emerging technology? This report represents a concrete first step toward meeting this
goal. We find that AI can be used to help address patterns of harm and thus reduce the
likelihood of harm. We then discuss some potential areas of focus militaries could prioritize
in order to reduce risks to civilians overall.
15. SUBJECT TERMS 15. 主题术语人工智能，AI，LOAC，自治，平民伤亡

Artificial Intelligence, AI, LOAC, autonomy, civilian casualties
16. SECURITY CLASSIFICATION OF: 17. LIMITATION 18. NUMBER 19a. NAME OF RESPONSIBLE PERSON 16. 安全分类: 17. 摘要的限制 18. 页数 19A.负责人姓名知识中心/Tanya McCants
OF ABSTRACT OF PAGES Knowledge Center/Tanya McCants
a. REPORT b. ABSTRACT c. THIS PAGE 19b. TELEPHONE NUMBER (include area a. REPORT b. ABSTRACT C.本页U 19B.电话号码(包括区号)703-824-2123
U U U code) U U
SAR 82 SAR 82
703-824-2123
Standard Form 298 (Rev. 8-98) 标准表格298(Rev.8-98)
Prescribed by ANSI Std. Z39.18 ANSI Std规定。 Z39.18
Executive Summary 行政摘要
Countries around the world have taken early steps to leverage artificial intelligence (AI) in 世界各国已经采取了早期措施，在军事能力中利用人工智能（AI）。尽管军方正在寻求利
military capabilities. Although militaries are seeking to leverage the technology of AI for 用人工智能技术提高效率和效率，但将人工智能应用于军事应用的想法也引起了相当大的争
greater effectiveness and efficiency, the idea of adapting AI to military applications has also 议。人们表达了许多担忧，包括潜在的偏见和缺乏公平性，以及在参与决策中保持人类的
判断力和责任。尽管如此，国际讨论的主要关注点是人工智能的军事应用是否可能本质上
created considerable controversy. Many concerns have been voiced, including potential bias 是不分青红皂白的，无法区分有效的军事目标和平民。
and lack of fairness, and maintaining human judgment and responsibility in engagement
decisions. That said, the chief concern in international discussions is whether military
applications of AI could be inherently indiscriminate, unable to differentiate between valid
military targets and civilians.
One way to answer this question is to look at specific military applications of AI, including 回答这个问题的一个方法是研究人工智能的具体军事应用，包括自主系统，并研究技术和
autonomous systems, and examine both technical and operational considerations for how risks 操作方面的考虑，以确定平民如何产生风险以及如何减轻风险。例如，在《联合国某些常
to civilians may arise and how they can be mitigated. For example, several presentations during 规武器公约》关于致命性自主武器系统会议期间的若干专题介绍中，列举了一些自治系统的
例子，这些自治系统可用于军事作战任务，其方式符合国际法，并不构成对平民的不分青红
the United Nations Convention on Certain Conventional Weapons meetings on lethal 皂白的危险。同样，CNA之前的一份报告（AI安全：行动计划）考虑了AI的一些额外的军
autonomous weapon systems featured examples of autonomous systems that could be used 事战斗应用，以及如何通过操作和技术缓解步骤将这些应用对平民的风险降至最低。 1
for military warfighting tasks in ways that complied with international law and did not
represent an indiscriminate hazard to civilians. Similarly, a previous CNA report (AI Safety: An
Action Plan) considered some additional military warfighting applications of AI and how risks
to civilians from those applications could be minimized through both operational and technical
mitigation steps.1
Those discussions, however, only address one half of the two-fold responsibilities for civilian 然而，这些讨论只涉及国际人道主义法中关于保护平民的双重责任的一半，即军队不应直
protection found in International Humanitarian Law—the negative responsibility that 接攻击平民的消极责任。军方采取一切可行预防措施保护平民免受伤害的肯定责任相对被
militaries should not direct attacks on civilians. The affirmative responsibility for militaries to 忽视。关于人工智能和自主性，各国不应该只问它们如何能够履行其消极责任，确保人工
智能应用在战争中不会不分青红皂白。他们还应该问：我们如何利用人工智能来保护平民
take all feasible precautions to protect civilians from harm has been relatively neglected. With 免受伤害？如何利用人工智能来减轻战争造成的痛苦、伤害和破坏？
regard to AI and autonomy, states should not only be asking how they can meet their negative
responsibilities of making sure that AI applications are not indiscriminate in warfare. They
should also be asking: How can we use AI to protect civilians from harm? And how can AI be
used to lessen the infliction of suffering, injury, and destruction of war?
This report represents a concrete first step toward answering these questions. We begin by 这份报告代表了回答这些问题的具体第一步。我们首先确定导致平民伤害的问题。如果我
framing the problems that lead to civilian harm. If we understand that AI is a tool for solving 们理解人工智能是一种解决问题的工具，那么在我们理解如何使用这个工具之前，我们需要
problems, before we understand how this tool can be used, we need to understand the 了解
1 Larry Lewis, AI Safety: An Action Plan for the Navy, CNA, DOP-2019-U-021957-1Rev, October 2019. 1Larry Lewis，AI安全：海军行动计划，CNA，DOP-2019-U-021957-1Rev，2019年10月。
CNA Occasional Paper | i CNA临时文件/i

problems to be solved. What problems need to be solved to better protect civilians or otherwise 待解决的问题。为了更好地保护平民或以其他方式促进国际人道法的人道原则，需要解决
promote IHL’s principle of humanity?. Although the imperative for avoiding civilian harm is 哪些问题？. 虽然避免平民伤害的必要性已得到普遍承认，但这种伤害如何发生的具体机制
universally acknowledged, the specific mechanisms for how such harm occurs have never been 从未得到详细描述。平民伤害是如何发生的？
characterized in detail. How does civilian harm occur?
After synthesizing our body of work on civilian harm—including analysis of several thousand 在综合了我们关于平民伤害的一系列工作--包括对军事行动造成的几千起现实世界平民伤
real-world incidents of civilian harm from military operations—we answer this question, 害事件的分析--之后，我们回答了这个问题，提出了一个说明平民伤害如何发生的框架。
presenting a framework illustrating how civilian harm occurs. We then discuss how civilian 然后，我们讨论如何减轻平民伤害，包括平民保护生命周期，这表明了减轻伤害的全面方法
。我们还讨论了可以采取的具体缓解措施的一些例子，以减少平民伤害，以显示为实现减
harm can be mitigated, including a civilian protection life cycle, which demonstrates a 轻平民伤害的目标可能采取的行动。
comprehensive approach to mitigating harm. We also discuss some examples of specific
mitigation steps that can be taken to reduce civilian harm to show the kinds of actions that are
possible for meeting the goal of civilian harm mitigation.
We then present a model approach for identifying opportunities where AI could be used to help 然后，我们提出了一个模型方法，以确定人工智能可以用来帮助解决平民伤害问题的机会
address the problem of civilian harm, using the civilian protection life cycle to illustrate ，利用平民保护生命周期来说明潜在的行动。我们发现人工智能应用在整个生命周期中有
potential actions. We find many opportunities for AI applications across the life cycle. This high 很多机会。这种大量的潜在应用不应该让我们感到惊讶，因为平民伤害问题可能被视为与
总体上更大的军事行动空间相关的行动，行为和政策的缩影。
volume of potential applications ought not surprise us because the problem of civilian harm
may be viewed as a microcosm of actions, behaviors, and policies associated with the much
larger military operational space overall.
We also discuss specific potential applications of AI that address risk factors we have observed 我们还讨论了人工智能的具体潜在应用，这些应用解决了我们在现实操作中观察到的风险
in real-world operations, leveraging techniques that currently exist and in many cases have 因素，利用目前存在的技术，并且在许多情况下已经应用于其虽然我们注意到没有任何解
already been applied to other problems. Although we note that no solution will eliminate the 决方案能够消除平民伤害问题-军事行动始终会对平民产生非零风险-AI可用于帮助解决我
们观察到的伤害模式并减少然后，我们讨论一些国家可以优先考虑的潜在重点领域，以减
problem of civilian harm—military operations will always have a non-zero risk to civilians— 少对平民的总体风险。
AI can be used to help address patterns of harm we observe and reduce the likelihood of harm.
We then discuss some potential areas of focus states could prioritize to reduce risks to civilians
overall.
For example, based on our analysis of particularly beneficial mitigation steps for reducing 例如，基于我们对特别有益的缓解措施的分析，以减少适用于人工智能应用的平民伤害，
harm to civilians that are amenable to AI applications, we suggest the following functions as 我们建议以下功能作为有希望的起点:
promising starting points:
• Alerting the presence of transient civilians by using object identification to • 通过使用物体识别自动监测目标区域周围的其他个人，并在检测到他们时发出警报

automatically monitor for additional individuals around the target area and send an ，提醒瞬态平民的存在。这项申请将提请行动部队注意这些个人,否则这些部队可
alert if they are detected. This application would bring these individuals to the 能会盯着目标而错过短暂的平民存在。
attention of operating forces that may otherwise fixate on the target and miss
transient civilian presence.
• Detecting a change from collateral damage estimate by finding differences • 通过发现用于确定附带损害估计的图像与支持参与的最近图像之间的差异来检测附
between imagery used to determine the collateral damage estimate and more recent 带损害估计的变化。此应用程序可以帮助识别操作部队可能无法识别的小细节，但
imagery taken in support of an engagement. This application can help identify little 可能是意外的平民存在的线索，例如建筑物附近的额外车辆。
details that operating forces might not recognize but that could be cues of
unanticipated civilian presence, such as additional vehicles near a building.
CNA Occasional Paper | ii 中央社临时文件/ii

• Alerting a potential miscorrelation by helping to identify that a miscorrelation has • 通过帮助确定发生了误关联，提醒潜在的误关联。例如，应用程序可以识别被跟
taken place. For example, applications could recognize that a vehicle being tracked is 踪的车辆与以前被跟踪的车辆不同，这表明威胁车辆和民用车辆之间发生了交换。
not the same one that was being tracked previously, showing that a swap has occurred
between a threat vehicle and a civilian vehicle.
• Recognizing protected symbols by using AI/machine learning methods to identify • 通过使用AI/机器学习方法识别指定受保护对象（例如红十字或红新月）的可接受
accepted symbols for designating protected objects (e.g., red cross or red crescent) 符号，并相应地向操作员或指挥链发出警报，从而识别受保护符号。受保护符号的
and alerting the operator or the chain of command accordingly. The presence of 存在并不意味着该地点实际上受到保护而不受攻击：该地点可能已经失去保护，或
者不择手段的一方可能违反国际法使用该符号来阻止攻击。但这种能力将提供一个
protected symbols does not mean that the location is, in fact, protected from attack: 安全网的情况下，受保护的符号存在，但被操作部队错过。 2
the location may have lost its protection or an unscrupulous party may be using the
symbol to deter attacks, in violation of international law. But this capability would
provide a safety net in cases where the protected symbol is present but was missed by
operating forces.2
Finally, by examining one tragic civilian harm incident in Afghanistan, we find we can draw 最后，通过审查阿富汗的一个悲惨的平民伤害事件，我们发现我们可以从我们的人工智能
from potential AI solutions from our AI applications matrix (including several of the 应用矩阵（包括上述几个应用）中借鉴潜在的人工智能解决方案，以帮助解决这一事件的根
abovementioned applications) to help address root causes in that incident and see how such 本原因，并看看这些解决方案如何有助于可能避免平民伤害。虽然这是我们研究结果的验
证步骤，但我们也注意到在这方面需要做更多的工作。这份报告只是探索细节非常重要的
solutions could help to possibly avert civilian harm. Although this is a validating step for our 广阔可能性空间的第一步。政府、军队和学术机构应该在这一基础上慎重地开发人工智能
findings, we also note that much more work needs to be done in this area. This report is merely 解决方案，以减轻对平民的伤害。
a first step in exploring a vast space of possibilities where details matter greatly. Governments,
militaries, and academic institutions should be deliberate in developing AI solutions to
mitigate harm to civilians, building on this foundation.
What remains is a matter of will, which we acknowledge is uncertain. Although militaries speak 剩下的是一个意志问题，我们承认这是不确定的。虽然军方谈到有助于减轻平民伤害的能
of capabilities that help mitigate civilian harm, such as precision-guided munitions, those 力，例如精确制导弹药，但这些能力是为了更有效地打击军事目标而获得的。虽然军队可
capabilities were acquired to engage military targets more effectively. Although militaries may 能具备在某些情况下有助于减轻伤害的能力，但军队并没有寻求—甚至没有认识到需要—从
我们在此确定的所有机制中全面发展减少平民风险的能力。因此，军队目前拥有的一套能
have capabilities that help to mitigate harm in some contexts, militaries have not sought—or 力是不完整的：还可以做更多的事情，现有的风险并不总是通过确实存在的能力来减轻。
even recognized the need—to comprehensively develop capabilities to reduce risk to civilians 例如，当平民被误认为军事目标并以这种错误的信念受到攻击时，精确制导弹药在减轻平民
from all the mechanisms we identify here. Therefore, the set of current capabilities held by 伤害方面没有价值。
militaries is incomplete: much more can be done, and existing risks are not always mitigated
by capabilities that do exist. For example, a precision-guided munition has no value in
mitigating civilian harm when civilians have been misidentified as a military target and are
attacked in that mistaken belief.
In summary, we do not observe militaries around the world seeking to field capabilities based 总之，我们没有看到世界各地的军队寻求基于其在减轻平民伤害方面的价值的野战能力。
on their value in mitigating civilian harm. We have taken a first step to show how AI-enabled 我们已经迈出了第一步，以展示人工智能和其他降低平民战争成本的应用是如何在可能的范
and other applications for reducing the cost of war on civilians are within the realm of the 围内实现的。军方是否会选择追捕他们还有待观察。
possible. It remains to be seen whether militaries will choose to pursue them.
2We note that the Australian Armed Forces have recognized this application as a promising one and have already 2我们注意到，澳大利亚武装部队已经认识到这个应用程序是一个有前途的应用程序，并且已经进行了实地实
conducted field experiments showing the utility of this function. 验，显示了这个功能的效用。
CNA Occasional Paper | iii 中央社临时文件/iii

This page is intentionally left blank. 本页有意留空.
CNA Occasional Paper | iv CNA临时文件/iv

Contents Contents
Introduction.................................................................................................................................................. 1 Introduction.................................................................................................................................................. 1
Concerns regarding civilian harm risks ..................................................................................................... 3 对平民伤害风险的担忧。.................................................................................................... 3
Opportunity for using AI to mitigate civilian harm ............................................................................... 4 利用人工智能减轻平民伤害的机会。.............................................................................. 4
Approach................................................................................................................................................................. 5 Approach................................................................................................................................................................. 5
Understanding Civilian Harm ................................................................................................................. 7 了解平民的伤害。................................................................................................................ 7
CNA’s work on civilian harm .......................................................................................................................... 7 CNA关于平民伤害的工作。......................................................................................................................... 7

Developing a methodology for understanding civilian harm ................................................... 7 制定了解平民伤害的方法。.................................................. 7
Applying the methodology: Iraq and Afghanistan ......................................................................... 8 应用方法：伊拉克和阿富汗。........................................................................ 8
Other assessments .................................................................................................................................... 12 其他评估。................................................................................................................................... 12
Root cause determinations: Why civilian harm happens ................................................................ 12 根本原因确定：为什么会发生平民伤害。............................................................... 12
A civilian protection life cycle for mitigating risks ............................................................................. 15 降低风险的平民保护生命周期。............................................................................ 15
Examples of practical steps for mitigating civilian harm................................................................. 17 减轻平民伤害的实际步骤的例子。................................................................ 17
Potential civilian harm mitigation measures for AI ........................................................................... 20 AI的潜在平民伤害缓解措施。.......................................................................... 20
Detecting transient civilian presence................................................................................................ 20 探测短暂的平民存在。............................................................................................... 20
Recognizing protected symbols........................................................................................................... 20 识别受保护的符号。.......................................................................................................... 20
Developing a robust civilian pattern of life .................................................................................... 21 发展稳健的平民生活模式。................................................................................... 21
Improving collateral damage estimates ........................................................................................... 21 改善附带损害估计。.......................................................................................... 21
Detecting misassociations ..................................................................................................................... 21 Detecting misassociations ..................................................................................................................... 21
Deconflicting with critical infrastructure ........................................................................................ 22 与关键基础设施解除冲突。....................................................................................... 22
Leveraging AI Applications for Mitigating Civilian Harm...........................................................23 利用人工智能应用程序减轻平民伤害。.......................................................... 23
Connecting the CPLC with AI/ML applications .................................................................................... 23 将CPLC与AI/ML应用程序连接。................................................................................... 23

Autonomous UASs/UAVs ....................................................................................................................... 25 Autonomous UASs/UAVs ....................................................................................................................... 25
Cyberspace ................................................................................................................................................... 27 Cyberspace ................................................................................................................................................... 27
Computer Vision ........................................................................................................................................ 29 计算机视觉。....................................................................................................................................... 29
Enabling Technologies ............................................................................................................................ 31 使能技术。........................................................................................................................... 31
Knowledge Representation and Reasoning.................................................................................... 31 知识表示和推理。................................................................................... 31
Natural Language Processing ............................................................................................................... 32 自然语言处理。.............................................................................................................. 32
Planning and Scheduling ........................................................................................................................ 33 计划和调度。....................................................................................................................... 33
Decision-Making ........................................................................................................................................ 34 Decision-Making ........................................................................................................................................ 34
Targeting ....................................................................................................................................................... 36 Targeting ....................................................................................................................................................... 36
Threat Monitoring ..................................................................................................................................... 38 威胁监控。.................................................................................................................................... 38
Training ......................................................................................................................................................... 38 Training ......................................................................................................................................................... 38
Commentary about the CPLC-AI matrix in Figure 3.................................................................... 40 关于图3中的CPLC-AI矩阵的评论。................................................................... 40
Reducing but not eliminating risks to civilians.................................................................................... 44 减少但不是消除对平民的风险。................................................................................... 44
CIVCAS Cause-Mitigations-AI-applications (CIVCAS-AI) matrix .................................................. 48 CIVCAS Cause-Mitigations-AI-applications (CIVCAS-AI) matrix .................................................. 48
A drone strike in Kabul: Could AI have helped? .................................................................................. 51 喀布尔的无人机袭击：AI可以提供帮助吗？ .................................................................................. 51
Conclusions .................................................................................................................................................52 Conclusions .................................................................................................................................................52
CNA Occasional Paper | v CNA临时文件/v

Appendix: AI/ML Approaches, Methods, and Algorithms Taxonomy ....................................54 附录：AI/ML方法、方法和算法分类。................................... 54
Figures ..........................................................................................................................................................59 Figures ..........................................................................................................................................................59

Abbreviations .............................................................................................................................................60 Abbreviations .............................................................................................................................................60
References ...................................................................................................................................................61 References ...................................................................................................................................................61
CNA Occasional Paper | vi CNA临时文件/vi

Introduction Introduction
The past decade has seen exponential progress in artificial intelligence (AI), defined as “the 在过去的十年里，人工智能（AI）取得了指数级的进步，人工智能被定义为"数字计算机或
ability of a digital computer or computer-controlled robot to perform tasks commonly 计算机控制的机器人执行通常与智能生物相关的任务的能力。"3人工智能正在对生活的许
associated with intelligent beings.”3 AI is having a transformative effect on many areas of life, 多领域产生变革性影响，包括商业、医学和银行业。示例包括以下内容:
including commerce, medicine, and banking. Examples include the following:
• Amazon’s logistics and delivery system, which uses machine learning (ML) AI to • 亚马逊的物流和交付系统，它使用机器学习（ML）AI来通知和优化这一过程。这
inform and optimize that process. This system includes an ML-driven decision on the 个系统包括一个ML驱动的决定最佳包装类型的数百万产品。 4
best packaging type for each of many millions of products.4
• Medical diagnoses using medical imagery can be performed faster and in a repeatable • 使用深度学习（DL）AI方法，可以更快，更经济地以可重复的方式执行使用医学图
and more economical way using deep learning (DL) AI methods.5 像的医学诊断。 5
• Banks use AI/ML models to detect suspicious behavior and stop fraudulent • 银行使用AI/ML模型来检测可疑行为并阻止欺诈交易，从而显着减少银行欺诈，这
transactions, leading to a significant reduction in banking fraud, the “biggest challenge 是金融业面临的"最大挑战"。” 6
for the financial industry.”6
Countries around the world have taken early steps to leverage AI in military capabilities, 世界各国已经采取了早期措施，在军事能力中利用人工智能，包括将人工智能用于自主系
including using AI for autonomous systems and functions, decision aids, and optimization in 统和功能、决策辅助和问题解决优化。例如，美国国防部（DOD）以战略，政策，组织变
革和资源的形式做出了严肃和公开的承诺，以利用人工智能。它还在开发利用现代人工智
problem solving. For example, the US Department of Defense (DOD) has made serious and
能技术的系统。示例包括以下内容:
public commitments in the form of strategy, policy, organizational changes, and resources to
leverage AI. It is also developing systems that leverage modern AI technologies. Examples
include the following:
3Encyclopedia Britannica, s.v. “artificial intelligence,” accessed Sept. 20, 2021, 3大英百科全书，s.v."人工智能，"访问9月。 2021年https://www.britannic
https://www.britannica.com/technology/artificial-intelligence. a.com/technology/artificial-intelligence ...
4 Amazon, “In the News: How Amazon Is Using Machine Learning to Eliminate 915,000 Tons of Packaging,” Jan. 29, 4亚马逊，"在新闻中：亚马逊如何使用机器学习来消除915,000吨包装，"1月1日。 2021年https://aws.amazon
2021, https://aws.amazon.com/blogs/industries/how-amazon-is-using-machine-learning-to-eliminate-915000- .com/blogs/industries/how-amazon-is-using-machine-learning-to-eliminate-915000tons-of-packaging/。
tons-of-packaging/.
5Ravi Aggarwal et al., “Diagnostic Accuracy of Deep Learning in Medical Imaging: A Systematic Review and Meta- 5Ravi Aggarwal等人。，"医学成像中深度学习的诊断准确性：系统评价和元分析，"npj数字医学4，第65条（
Analysis,” npj Digital Medicine 4, article 65 (Apr. 7, 2021), https://doi.org/10.1038/s41746-021-00438-z. Apr。 7，2021），https://doi.org/10.1038/s41746-021-00438-z ...
6Robin Trehan, “How AI Is Transforming Fraud Prevention in Banking and Finance,” Deltec, May 11, 2020, 6Robin Trehan，"AI如何改变银行和金融领域的欺诈预防"，Deltec，2020年5月11日，https://www.delte
https://www.deltecbank.com/2020/05/11/how-ai-is-transforming-risk-in-finance-and-banking/?locale=en. cbank.com/2020/05/11/how-ai-is-transforming-risk-in-finance-and-banking/?locale=en ...
1 1
CNA Occasional Paper | 中央社临时文件|
• In 2017, DOD launched Project Maven, an ongoing effort to tap AI to help DOD analyze • 2017年，国防部启动了Maven项目，这是一项正在进行的利用人工智能的努力，
what was fast becoming an overwhelming amount of full-motion video being collected 以帮助国防部分析为支持反暴乱和反恐行动而收集的大量全动态视频。 7
in support of counterinsurgency and counterterrorism operations.7
• DOD is conducting an initiative to leverage AI to dramatically speed up assessments • 国防部正在开展一项利用人工智能的举措，以显着加快援助人道主义和救灾任务所
needed to aid in humanitarian and disaster relief missions, including route analysis, 需的评估，包括路线分析、损害评估、洪水探测和火灾周边分析。这项努力得到了
damage assessment, flood water detection, and fire perimeter analysis. This effort has 约翰霍普金斯大学应用物理实验室的支持，并且应用程序已经过国家国民警卫队单
位的测试。 8
been supported by Johns Hopkins Applied Physics Laboratory, and applications have
been tested by state National Guard units.8
• DOD is exploring the use of small autonomous, AI-enabled unmanned aerial vehicles • 国防部正在探索使用小型自动驾驶，人工智能无人驾驶飞行器（Uav）由特种作战
(UAVs) for operation by special operations forces in indoor environments.9 部队在室内环境中操作。 9
• In 2020, the US Air Force demonstrated the use of AI on board a US military aircraft • 在2020中，美国空军首次在美国军用飞机上展示了人工智能的使用。这架载人飞
for what appears to be the first time. The manned aircraft—a U-2 Dragon Lady with 机—一架带有机载飞行员的U-2龙夫人--使用人工智能算法来控制飞机的传感器和
an onboard pilot—used an AI algorithm to control the aircraft’s sensor and navigation 导航系统。人工智能技术（不与飞机的飞行控制交互）是专门设计的，没有手动复
盖。 10
systems. The AI technology (which did not interact with the aircraft’s flight controls)
was specifically designed without a manual override.10
The US is only one of many countries making AI the center of its strategy. China, another world 美国只是众多将人工智能作为其战略中心的国家之一。中国是人工智能发展的另一个世界
leader in the development of AI, is doing the same and pursuing AI for a wide range of national 领导者，也在做同样的事情，并在广泛的国家应用中追求人工智能，包括军事应用。示例
applications, including military applications. Examples include the following: 包括以下内容:
• AI for target identification, including identifying US military aircraft on runways • 用于目标识别的AI，包括识别跑道上的美国军用飞机（报告准确率为92％）和海上

(reporting an accuracy of 92 percent) and naval targets in maritime environments11 环境中的海军目标11
• AI-enabled capabilities in facial recognition, (Chinese language) textual analysis, and • 人工智能在面部识别、中文文本分析以及大量监控数据的分析和综合方面的能力。
the analysis and synthesis of massive troves of surveillance data, which the country
7 Cheryl Pellerin, “Project Maven to Deploy Computer Algorithms to War Zone by Year’s End,” DOD website, July 7Cheryl Pellerin，"Project Maven To Deploy Computer Algorithms to War Zone by Year'S End，"国防部网站
21, 2017, https://www.defense.gov/Explore/News/Article/Article/1254719/project-maven-to-deploy- ，2017年7月21日，https://www.defense.gov/Explore/News/Article/Article/1254719/project-maven-to-d
computer-algorithms-to-war-zone-by-years-end/. eploycomputer-algorithms-to-war-zone-by-years-end/。
8 Luke Strum, “Intel Airmen Sharpen AI Technology for Domestic Response,” Air National Guard website, Nov. 25, 8Luke Strum，"英特尔飞行员为国内反应锐化AI技术"，Air National Guard网站，11月11日。 2019年2月25日
2019, https://www.ang.af.mil/Media/Article-Display/Article/2025332/intel-airmen-sharpen-ai-%20technology- https://www.ang.af.mil/Media/Article-Display/Article/2025332/intel-airmen-sharpen-ai-%20technologyfor-
for-domestic-response/; Department of Defense, “Mission Initiatives: Threat Reduction and Protection,” JAIC domestic-response/;国防部,"使命倡议:减少威胁和保护,"JAIC网站,https://www.ai.mil/mi_threat_reduction_an
website, https://www.ai.mil/mi_threat_reduction_and_protection.html. d_protection.html ...
9Department of Defense, “Mission Initiatives: Joint Warfighting Operations,” JAIC website, 9国防部，"任务倡议：联合作战行动"，JAIC网站，https://www.ai.mil/mi_joint_warfig
https://www.ai.mil/mi_joint_warfighting_operations.html. hting_operations.html ...
10Aaron Gregg, “In a First, Air Force Uses AI on Military Jet,” Washington Post, Dec. 16, 2020, 10Aaron Gregg，"首先，空军在军用喷气式飞机上使用人工智能，"华盛顿邮报，12月10日。 2020
年16月https://www.washingtonpost.com/business/2020/12/16/air-force-artificial-intelligence/。
https://www.washingtonpost.com/business/2020/12/16/air-force-artificial-intelligence/.
11Alex Barker, “Giving Precision Munitions ‘Eyes’ and a ‘Brain’: The State of PLA Research on Military Target 11亚历克斯*巴克，"赋予精确弹药'眼睛'和'大脑'：解放军对军事目标识别的研究现状"，中国简报，21，第1
Recognition,” China Brief, 21, no. 13 (Jul. 2, 2021), https://jamestown.org/program/giving-precision-munitions- 3期。 2，2021年），https://jamestown.org/program/giving-precision-munitionseyes-and-a-brain-the-s
eyes-and-a-brain-the-state-of-pla-research-on-military-target-recognition/. tate-of-pla-research-on-military-target-recognition/。
2 2
has been able to develop and refine through extensive surveillance of its domestic 已经能够通过广泛监测其国内人口为国内目的而发展和完善12
population for domestic purposes12
• AI-enabled unmanned systems, including patrol boats and swarms of armed drones13 • 支持AI的无人系统，包括巡逻艇和成群的武装无人机13
Similarly, Russia is making AI a major priority for modernization of its forces, seeing innovation 同样，俄罗斯正在将人工智能作为其部队现代化的主要优先事项，将创新视为其国防和大
as essential to its defense and status as a great power. Specific efforts include the following: 国地位的关键。具体努力包括以下内容:
• Experimentation with unmanned, AI-enabled small tanks and robotic exoskeletons for • 为士兵进行无人驾驶，人工智能启用的小型坦克和机器人外骨骼的实验，这将使部
soldiers that will allow troops to carry more weapons and equipment14 队能够携带更多的武器和装备14
• Using AI to automate processes and command and control functions such as collection, • 使用人工智能自动执行流程和指挥和控制功能，例如收集、处理、存储和交付优化
processing, storage, and delivery of information necessary to optimize command and 部队和武器指挥和控制所需的信息
control of troops and weapons
• Employing AI on the Su-35S, a heavy long-range fighter, to support pilot decision- • 在重型远程战斗机Su-35s上使用AI来支持飞行员的目标获取和战斗机动决策15
making for target acquisition and combat maneuver15
These are only a few of the military AI applications being explored by many countries around 这些只是世界上许多国家正在探索的军事人工智能应用中的一小部分。
the world.
Concerns regarding civilian harm risks 对平民伤害风险的关注

Although many militaries are seeking to leverage the technology of AI, the idea of adapting AI 尽管许多军队正在寻求利用人工智能技术，但将人工智能适应军事应用的想法也引起了相
to military applications has also created considerable controversy. The most deliberate debate 当大的争议。关于这个问题的最深思熟虑的辩论涉及致命的自主武器系统（法律）。基于
所表达的关切，2014年，联合国（UN）《特定常规武器公约》（CCW）首次就法律进行非
on this issue concerns lethal autonomous weapon systems (LAWS). Based on expressed
正式讨论。这个机构现在已经花了近十年的时间来讨论法律的道德，法律和操作考虑，包
concerns, in 2014 the United Nations (UN) Convention on Certain Conventional Weapons 括是否应该允许自主运行的武器系统（没有人类操作员）
(CCW) first met for informal discussions on LAWS. This body has now spent almost a decade
discussing the ethical, legal, and operational considerations of LAWS, including whether
weapon systems operating autonomously (without a human operator) should be allowed to
12C. Todd Lopez, “Where It Counts, U.S. Leads in Artificial Intelligence,” DOD website, July 9, 2020, 12C.Todd Lopez,"Where It Counts,U.S.Leads in Artificial Intelligence,"国防部网站，2020年7月9日，https:
https://www.defense.gov/Explore/News/Article/Article/2269200/where-it-counts-us-leads-in-artificial- //www.defense.gov/Explore/News/Article/Article/2269200/where-it-counts-us-leads-in-artificialintelli
intelligence/; Yasmin Tadjdeh, “China Threatens U.S. Primacy in Artificial Intelligence (UPDATED),” National gence/;Yasmin Tadjdeh，"中国威胁美国在人工智能中的首要地位（更新），"国防，10月10日。 2020年3
Defense, Oct. 31, 2020, https://www.nationaldefensemagazine.org/articles/2020/10/30/china-threatens-us- 1月https://www.nationaldefensemagazine.org/articles/2020/10/30/china-threatens-usprimacy-in-artifi
primacy-in-artificial-intelligence. cial-intelligence ...
13 Yasmin Tadjdeh, “China Threatens U.S. Primacy.” 13Yasmin Tadjdeh，"中国威胁美国的首要地位。”
14Margarita Konaev and Samuel Bendett, “Russian AI-Enabled Combat: Coming to a City near You?” War on the 14Margarita Konaev和Samuel Bendett，"俄罗斯AI启用的战斗：来到你附近的城市？"岩石上的战争，2019
Rocks, July 31, 2019, https://warontherocks.com/2019/07/russian-ai-enabled-combat-coming-to-a-city-near- 年7月31日，https://warontherocks.com/2019/07/russian-ai-enabled-combat-coming-to-a-city-nearyo
you/. u/。
15Jeffrey Edmonds et al., Artificial Intelligence and Autonomy in Russia, CNA, DRM-2021-U-029303-Final, May 15Jeffrey Edmonds等人。，俄罗斯的人工智能和自治，CNA，DRM-2021-U-029303-Final，2021年5月
2021. 。
3 3
use lethal force.16 Although autonomy is the AI application most debated in the international 使用致命武力。 16尽管自治是国际社会争论最多的人工智能应用但人们越来越意识到其他
community, there is growing awareness that other applications—decision aids and optimized 应用—决策辅助和优化功能-也可能带来风险。
functions—can also carry risk.
What are the risks involved? Many concerns have been voiced, some valid (e.g., potential bias 涉及哪些风险？人们表达了许多担忧，其中一些是有效的（例如，潜在的偏见和缺乏公平
and lack of fairness, maintaining human judgment and responsibility in engagement decisions) 性，在参与决策中保持人类的判断和责任），而一些则不太有效（例如，对机器人接管世
and some less valid (e.g., the fear of robots taking over the world) for the current state of AI 然而，国际讨论的主要关切是，人工智能的军事应用是否可能本质上是不分青红皂白的，无
法区分有效的军事目标和平民。
technology. However, the chief concern in international discussions is whether military
applications of AI could be inherently indiscriminate, unable to differentiate between valid
military targets and civilians.
One way to answer this question is to look at specific military applications of AI, including 回答这个问题的一个方法是研究人工智能的具体军事应用，包括自主系统，并研究技术和
autonomous systems, and examine both technical and operational considerations for how risks 操作方面的考虑，以确定平民如何产生风险以及如何减轻风险。在联合国《特定常规武器
to civilians may arise and how they can be mitigated. Several presentations during the UN CCW 公约》会议上的一些发言确实列举了自治系统的例子，这些系统可以用于军事作战任务，其
方式符合国际法，不会对平民造成不分青红皂白的危害。同样，CNA之前的一份报告考虑
meetings did indeed feature examples of autonomous systems that could be used for military 了人工智能的一些额外的军事战争应用，以及如何通过操作和技术缓解步骤将这些应用对平
warfighting tasks in ways that complied with international law and did not represent an 民的风险降至最低。 17
indiscriminate hazard to civilians. Similarly, a previous CNA report considered some additional
military warfighting applications of AI and how risks to civilians from those applications could
be minimized through both operational and technical mitigation steps.17
Opportunity for using AI to mitigate civilian 利用人工智能减轻平民伤害的机会

harm
Another question should be asked because of the nature of international humanitarian law 由于国际人道主义法的性质，应该提出另一个问题。各国制定了国际人道法，使其及其武
(IHL). States established IHL to legally obligate them and their armed forces to standards of 装部队在法律上有义务遵守武装冲突中的行为标准，特别强调保护平民。所有国家都自愿
conduct in armed conflict, with particular emphasis on the protection of civilians. All states 承担这些义务，因为它们认识到这样做的道义和战略重要性。国际人道法既包括肯定责任
（例如，军队应采取一切可行的预防措施保护平民免受伤害），也包括消极责任（例如，军
took on these obligations willingly because they recognized the moral and strategic importance 队不应直接攻击平民）。国际人道法还申明各国承诺：
of doing so. IHL includes both affirmative responsibilities (e.g., militaries should take all
feasible precautions to protect civilians from harm) and negative responsibilities (e.g.,
militaries should not direct attacks on civilians). IHL also affirms the commitment of states to
16 The CCW is properly referred to as the Convention on Prohibitions or Restrictions on the Use of Certain 16《特定常规武器公约》被恰当地称为《禁止或限制使用某些可被认为具有过分伤害力或滥杀滥伤作用的
Conventional Weapons Which May Be Deemed to Be Excessively Injurious or to Have Indiscriminate Effects. 常规武器公约》。
17 Larry Lewis, AI Safety: An Action Plan for the Navy, CNA, DOP-2019-U-021957-1Rev, October 2019. 17Larry Lewis，AI安全：海军行动计划，CNA，DOP-2019-U-021957-1Rev，2019年10月。
4 4
the principle of humanity, which “forbids the infliction of all suffering, injury or destruction not 人道原则，该原则"禁止为实现冲突的合法目的而施加一切不必要的痛苦、伤害或破坏。” 1
necessary for achieving the legitimate purpose of a conflict.”18 8
Thus, regarding AI, including autonomous functions, although states should be asking how to 因此，关于人工智能，包括自主功能，虽然各国应该问如何履行确保人工智能应用不滥杀
meet their negative responsibilities of ensuring that AI applications are not indiscriminate, 滥伤的消极责任，但它们也应该问如何更好地履行自己确定的积极义务，即：我们如何利用
they should also be asking how to better fulfill their self-determined positive obligations, 人工智能保护平民不受伤害？我们怎样才能减少整体上的痛苦、伤害和破坏呢？
namely: How can we use AI to protect civilians from harm? And how can we lessen the infliction
of suffering, injury, and destruction overall?
We note that CNA is not alone in making this observation. After we began this project, Australia, 我们注意到，CNA并不是唯一一个进行这一观察的人。在我们开始这个项目之后，澳大利
Canada, Japan, South Korea, the United Kingdom (UK), and the United States submitted a joint 亚，加拿大，日本，韩国，英国（英国）和美国向联合国《特定常规武器公约》提交了一份
联合文件，其中包括寻求这些问题答案的具体而言，该文件建议致命自主武器系统政府专
paper to the UN CCW that included a similar recommendation to seek answers to these
家组在2022年将重点转移到包括"确定法律领域新兴技术可用于减少军事行动中平民风险的
questions. Specifically, the paper recommended the Group of Government Experts on Lethal 方法示例。"19本报告是实现这一目标的具体第一步。
Autonomous Weapon Systems shift its focus in 2022 to include “identifying examples of ways
in which emerging technologies in the area of LAWS could be used to reduce the risks to
civilians in military operations.”19 This report represents a concrete first step in meeting this
goal.
Approach Approach
If we understand that AI is a tool for solving problems, before we understand how this tool can 如果我们理解AI是解决问题的工具，在我们理解这个工具如何使用之前，我们需要了解要
be used, we need to understand the problems to be solved. What are the problems that need 解决的问题。为了更好地保护平民或以其他方式促进国际人道法的人道原则，需要解决哪
些问题？虽然避免平民伤害的必要性已得到普遍承认，但这种伤害如何发生的具体机制从
to be solved to better protect civilians or otherwise promote IHL’s principle of humanity?
未得到详细描述。平民伤害是如何发生的？在综合了我们关于平民伤害的一系列工作--包
Although the imperative for avoiding civilian harm is universally acknowledged, the specific 括对军事行动造成的数千起现实世界平民伤害事件的分析--之后，我们为平民伤害如何发
mechanisms for how such harm occurs have never been characterized in detail. How does 生制定了一个框架。我们还讨论了最近美国和联盟行动中观察到的一些特定风险因素。
civilian harm occur? After synthesizing our body of work on civilian harm—including analysis
of several thousand real-world incidents of civilian harm from military operations—we
produce a framework for how civilian harm occurs. We also discuss some specific risk factors
observed in recent US and coalition operations.
We then discuss some examples of specific mitigation steps that can be taken to reduce civilian 然后，我们讨论了可以采取的具体缓解措施的一些例子，以减少平民伤害，以表明可以采
harm to show what kinds of actions are possible for meeting the goal of civilian harm 取哪些行动来实现平民伤害的目标
18International Committee of the Red Cross, International Humanitarian Law: Answers to Your Questions, June 9, 18红十字国际委员会，《国际人道主义法：对你的问题的回答》，2020年6月9日。国际人道法的这一原则部
2020. This principle of IHL stems in part from the Martens Clause, discussed here: Theodor Meron, “The Martens 分源于马滕斯条款，这里讨论过：Theodor Meron，"马滕斯条款，人道原则和公众良心的支配"，美国国际法
Clause, Principles of Humanity, and Dictates of Public Conscience,” The American Journal of International Law 94, 杂志94，第1期（1月1日）。 2000).
no. 1 (Jan. 2000).
19 Australia, Canada, Japan, South Korea, the United Kingdom, and the United States, Discussion Paper – Building on 19澳大利亚、加拿大、日本、韩国、联合王国和美国2021年6月提交给联合国《特定常规武器公约》(《特定常
Chile’s Proposed Four Elements of Further Work for the Convention on Certain Conventional Weapons (CCW) Group 规武器公约》)关于致命自主武器系统(法律)领域新兴技术的政府专家组(GGE)拟议进一步工作的四个要素的讨论
of Governmental Experts (GGE) on Emerging Technologies in the Area of Lethal Autonomous Weapons Systems 文件。
(LAWS), submitted to the UN CCW June 2021.
5 5
mitigation. We then present a framework for how AI could be used to help address the problem 缓解。然后，我们提出了一个如何利用人工智能来帮助解决平民伤害问题的框架。最后，
of civilian harm. Finally, we discuss some specific applications of AI that address some risk 我们讨论了人工智能的一些具体应用，这些应用解决了我们观察到的一些风险因素，说明了
factors we have observed, illustrating some potential areas of focus states could prioritize to 各国可以优先考虑的一些潜在重点领域，以减少对平民的总体风险。
reduce risks to civilians overall.
6 6
Understanding Civilian Harm 了解平民危害
The tragedy of civilian harm has always been a feature of warfare, and over time states have 平民伤害的悲剧一直是战争的一个特点，随着时间的推移，各国试图限制这种伤害，包括
sought to limit this harm, including the development of IHL. However, over the past several 国际人道法的发展。然而，在过去几十年中，我们对平民伤害有了更好的了解，并对如何
decades, we have come to a better understanding of civilian harm and a more complete view 减轻这种伤害有了更全面的认识。中心一直处于这一努力的最前沿,通过分析现实世界的平
民伤害事件,将科学方法应用于平民伤害问题。我们在本章开始时描述了CNA关于平民伤害
of how to mitigate this harm. CNA has been at the forefront of this effort, applying the scientific 的工作，然后我们讨论了对造成平民伤害的机制的关键见解，包括理解这些原因的框架。
method to the problem of civilian harm through analysis of real-world civilian harm incidents. 最后，根据我们确定的原因，我们讨论了减轻平民伤害的具体解决方案，这可以作为考虑潜
We begin this chapter by describing CNA’s body of work on civilian harm, and then we discuss 在人工智能应用的起点。
key insights into the mechanisms that cause civilian harm, including a framework for
understanding these causes. Finally, based on our identified causes, we discuss specific
solutions for mitigating civilian harm, which can serve as a starting point for considering
potential AI applications.
CNA’s work on civilian harm 中央社关于平民伤害的工作

Our work on civilian harm began through investigation of a different, but related, operational 我们关于平民伤害的工作是通过调查一个不同但相关的行动问题开始的：友好火力。在19
problem: friendly fire. In Operation Desert Storm in 1991, a significant fraction of US casualties 91年的沙漠风暴行动中，美国很大一部分伤亡是由友军火力造成的，美军在美军交战中丧
were caused by friendly fire, where US forces were killed or wounded by US engagements. 生或受伤。国防部认识到友军火力是一个重大问题，因此创建了联合防空行动联合交战区
联合测试和评估活动，以努力发展能力和战术，帮助美军在共同的战斗空间中更好、更安全
Recognizing friendly fire to be a significant problem, DOD created the Joint Air Defense 地协同作战。这项活动得到了持续，并改名为全服战斗识别评估小组，后来又改名为联合
Operations Joint Engagement Zone Joint Test and Evaluation Activity in an effort to develop 战斗识别
capabilities and tactics to help US forces to operate together better and more safely in a
common battlespace. This activity was sustained and renamed the All-Service Combat
Identification Evaluation Team and later renamed again as the Joint Combat Identification
Evaluation Team, with activities including a wider variety of missions and domains.20 评估小组，其活动包括更广泛的任务和领域。 20
Developing a methodology for understanding civilian harm 制定了解平民伤害的方法

CNA supported these activities by examining challenges in combat identification, including CNA通过审查战斗识别方面的挑战来支持这些活动，包括重建和纵向分析在多个现场事件
reconstruction and longitudinal analysis of constructive friendly fire incidents occurring in 中发生的建设性友好火灾事件。这种方法产生了一些以前没有确定的发现。当美国在2003
multiple live events. This methodology yielded several findings that had not previously been 开始伊拉克自由行动时，CNA对实际友好火灾事件采用了相同的方法，使用所有可用信息，
包括调查报告，操作数据（包括来自战斗系统和数据链此外，一个
identified. When the US began Operation Iraqi Freedom in 2003, CNA employed the same
methodology for actual friendly fire incidents, using all available information, including
investigation reports, operational data (including digital data from combat systems and data
links), and media reports to give insights about the root causes of friendly fire. In addition, one
20 Larry Lewis, Insights for the Third Offset, CNA, DRM-2017-U-016281-Final, Sept. 2017. 20Larry Lewis,Insights for The Third Offset,CNA,DRM-2017-U-016281-Final,Sept. 2017.
7 7
of the friendly fire incidents resulted in civilian harm because civilians were in the proximity 在友军火灾事件中，平民受到伤害，因为平民靠近友军。 CNA对这一事件的详细重建导致
of friendly forces. CNA’s detailed reconstruction of this incident resulted in the identification 确定了几个以前不了解的因果因素。例如，这一事件的特点是：飞行员错误地将有关威胁
of several causal factors that were not understood before. For example, the incident featured a 的信息与平民和友军的位置联系起来，导致飞行员接触错误的位置，造成平民和友军伤亡。
这些新的见解表明了我们的方法对于理解和解决平民伤害问题的价值。
misassociation: the pilot wrongly associated information regarding the threat with the location
of civilians and friendly forces, leading the pilot to engage the wrong location and cause civilian
and friendly casualties. These new insights suggested our methodology’s value for
understanding and addressing the problem of civilian harm.
Applying the methodology: Iraq and Afghanistan 应用方法：伊拉克和阿富汗

As insurgencies developed in Iraq and Afghanistan (Operation Enduring Freedom, launched in 随着伊拉克和阿富汗叛乱的发展（2001年发起的持久自由行动），美国从主要的战斗行动
2001), the US shifted from major combat operations to counterinsurgency, an approach for 转向了平叛，这一方针在很大程度上没有做好准备。由于平民保护是反暴动的一个核心特
which it was largely unprepared. With civilian protection as a central feature of 征,而且对不穿制服或其他识别资料的非正常威胁的军事人员来说,身份查验的挑战也增加了,
因此平民的伤害成为一个核心挑战。 CNA分析了两场冲突中的平民伤害，详情如下。
counterinsurgency and the added identification challenges for a military with an irregular
threat that does not wear uniforms or other identifying information, civilian harm became a
central challenge. CNA analyzed civilian harm in both conflicts, as detailed below.
Iraq: Analysis of escalation of force 伊拉克：对武力升级的分析

We first examined civilian harm in Iraq, where civilian casualties (civilian harm) were 我们首先检查了伊拉克的平民伤害，那里的平民伤亡（平民伤害）主要是由部队事件升级
primarily caused by escalation of force incidents. These incidents occurred both at military 造成的。这些事件既发生在军事检查站,也发生在车队行动期间,在这些车辆和个人在他们看
checkpoints and during convoy operations, where military forces engaged vehicles and 来受到威胁。这一原因造成的平民伤亡急剧上升促使首次以军事为主导的对平民伤害的跟
踪：多国部队-伊拉克于2004年开始跟踪平民伤害，以了解问题的范围并试图追踪显示，
individuals who appeared to them to be threatening. A sharp rise in civilian casualties from 美军在2005年上半年因武力事件升级造成500多名平民伤害。 2005年年中，在高级领导人
this cause prompted the first military-led tracking of civilian harm: Multinational Forces-Iraq 指示减轻平民伤害的推动下，美国军队改变了战术和程序，以减少对平民的风险，导致武力
began tracking civilian harm in 2004 to understand the scope of the issue and try to identify 升级造成的平民伤害显着下降。在这一下降之后，中央社分析了这一平民伤害数据，并评
solutions. Tracking indicated that US forces caused more than 500 civilian harm from 估了与部队事件升级有关的风险因素，例如错误识别的风险和在远超过自卫考虑所需的距离
escalation of force incidents in the first half of 2005. In mid-2005, prompted by senior leader 进行感知威胁的倾向--不幸的是，这些风险因素在阿富汗会再次出现。 21
direction to mitigate civilian harm, US forces changed their tactics and procedures to reduce
risks to civilians, resulting in a significant drop in civilian harm from escalation of force. After
this drop, CNA analyzed this civilian harm data and assessed risk factors associated with
escalation of force incidents, such as the risk of misidentification and the tendency to engage
perceived threats at distances far exceeding those needed for self-defense considerations—
risk factors that unfortunately would be seen again in Afghanistan.21
Afghanistan: Real-time support Afghanistan: Real-time support

The problem of civilian harm also became a strategic issue in US and international force 平民伤害问题也成为美国和国际部队在阿富汗行动中的一个战略问题。阿富汗领导人和国
operations in Afghanistan. Afghan leaders and the international community expressed alarm 际社会对美国和联盟在阿富汗造成平民伤害的人数不断增加表示震惊
over escalating numbers of US- and coalition-caused civilian harmcivilian harm in Afghanistan
21 Larry Lewis, Reducing and Mitigating Civilian Casualties: Enduring Lessons, Joint Staff, Apr. 12, 2013. 21Larry Lewis，减少和减轻平民伤亡：持久的教训，联合工作人员，Apr。 12, 2013.
8 8
between 2006 and 2009, and US leaders saw the issue of civilian harmcivilian harm becoming 在2006和2009之间，美国领导人看到平民危害问题成为运动中的一个战略问题。例如，
a strategic issue in the campaign. For example, the then-commander of US Central Command 当时的美国中央司令部指挥官将平民伤害描述为一个"有毒"问题，威胁到美国和国际战略目
described civilian harm as becoming a “toxic” issue that threatened the meeting of US and 标的实现，并加强了反政府分子的支持。
international strategic objectives and strengthened the support of antigovernment elements.
This realization drove a practical approach to the problem of civilian harm in Afghanistan. Like 这一认识推动了对阿富汗平民伤害问题的实际处理。就像在伊拉克一样，这个过程始于数
in Iraq, the process began with data: in late 2008, US and international forces began tracking 据：在2008后期，美国和国际部队开始追踪平民伤害。这项追踪工作最初是为了对付公共
civilian harm. This tracking effort was originally intended to counter external allegations for 事务方面的外部指控,但收集这些数据的过程成为采取更以证据为基础的办法减轻平民伤害
的基础。在2009中，由CNA领导的美国军方lessonslearned组织努力分析了行动中平民伤
the purpose of public affairs, but the process of gathering this data became a foundation for a 害事件。我们使用行动记录，并在可用时结合开源和民间社会信息，以发现平民伤害的模
more evidence-based approach to mitigating civilian harm. In 2009, a US military lessons- 式和机制。这一分析产生了新的见解，使军队能够采取切实、有重点的步骤，以减少在阿
learned organization effort led by CNA analyzed civilian harm incidents in the midst of 富汗行动中对平民造成的风险。根据这一分析所作的修改包括修改了国际安全援助部队(安
operations. We used operational records, combined with open source and civil society 援部队)关于进行空袭的战术指示,以及关于检查站部队升级的新指导,以及许多其他措施。
information when available, to discover patterns and mechanisms of civilian harm. This
analysis yielded new insights that enabled practical, focused steps military forces could take to
reduce the risk to civilians in operations in Afghanistan. Changes made based on this analysis
included a modified International Security Assistance Force (ISAF) tactical directive governing
the conduct of air strikes and new guidance for escalation of force at checkpoints, among many
other measures.
Subsequent analysis showed that these steps were effective in reducing civilian harmcivilian 随后的分析表明，这些步骤在减少平民危害方面是有效的，第一年内减少了20％，此后继
harm, with a reduction of 20 percent within the first year and continued reductions thereafter. 续减少。每个行动的平民伤害率代表行动期间平民伤害的相对风险，在可以衡量这种比率
The rate of civilian harmcivilian harm per operation, representing the relative risk of civilian 时也会下降。因此，与这些变化之前相比，行动不太可能造成平民伤害。这种减轻平民伤
害的实际办法包括(1)跟踪平民伤害,(2)监测风险并加以调整以减轻这些风险。 CNA在这两个
harm during operations, also decreased when such rates could be measured. Consequently, 步骤中都支持美国和国际部队。
operations were less likely to cause civilian harmcivilian harm than they were before these
changes. This practical approach to civilian harm mitigation included (1) tracking civilian
harm, and (2) monitoring risks and adapting to mitigate them. CNA supported US and
international forces in both these steps.
Tracking civilian harmcivilian harm 追踪平民危害

We noted above that the US began tracking incidents of civilian harmcivilian harm in Iraq in 我们在上面注意到，美国开始跟踪伊拉克平民伤害事件，以回应越来越多的关注，美国军
response to growing attention to civilians being killed by US forces at checkpoints through 队通过武力升级在检查站杀害平民。然而，这种追踪是一项临时措施，侧重于一种特定类
escalation of force. However, this tracking was a temporary measure that focused on one 型的行动：检查点。在阿富汗，安援部队开始全面追踪各类行动的平民伤害事件。在2009
中，跟踪被资源化以使其更强大，并且它开始用于促进操作学习。跟踪包括一个电子表格
particular type of operation: checkpoints. In Afghanistan, ISAF began tracking civilian ，其中包括日期、时间、事件地点、所涉单位、行动类型、平民伤害数量以及可用于后果管
harmcivilian harm incidents comprehensively for all types of operations. In 2009, the tracking 理和趋势分析的其他细节。这种跟踪成为后来美国冲突中复制的最佳实践。例如，反恐部
was resourced to make it more robust, and it started being used to promote operational 队和美国领导的反ISIS联盟都追踪其行动对平民的伤害。
learning. Tracking consisted of a spreadsheet with date, time, location of incident, unit
involved, type of operation, number of civilian harm, and other details that could be used for
consequence management and trend analysis. This tracking became a best practice replicated
in later US conflicts. For example, counterterrorism forces and the US-led counter-ISIS coalition
both tracked civilian harm from their operations.
9 9
In addition to aiding in consequence management (e.g., determining whether medical care 除了协助后果管理（例如，确定是否需要提供医疗护理，美国是否应该为疏忽的平民通行
needed to be provided, whether the US should apologize for inadvertent civilian tolls, or 费道歉，或者补偿等补偿是否适当）和通信（例如，收集有关特定事件对平民伤害的跟踪
whether amends such as compensation were appropriate) and communications (e.g., 绝不是完美的（正如我们稍后将讨论的那样），但它仍然为学习和适应提供了基础。中心
定期审查平民伤害追踪小组的数据库，其中载有个别事件的详细资料，并作出更正和改进，
gathering the latest information regarding civilian tolls from a particular incident), this 以提高作为缓解工作基础的资料的准确性。
tracking created a way to factor in contributions from external organizations and individuals
with relevant information. Tracking of civilian harm was by no means perfect (as we will
discuss later), but it still provided the foundation for learning and adaptation. CNA periodically
reviewed the civilian harm tracking cell’s database containing details of individual incidents
and made corrections and improvements to improve the accuracy of the information that
served as a foundation for mitigation efforts.
Monitoring and responding to trends 监测和应对趋势

A second key aspect to addressing the risk to civilians is monitoring metrics that capture that 解决平民风险的第二个关键方面是监测捕捉这种风险并对其作出反应的指标。除平民伤害
risk and responding to them. In addition to the number of civilian harm, another important 数外，另一项重要措施是平民伤害率--造成平民伤害的罢工次数除以罢工总数。这项措施
measure is the rate of civilian harm—the number of strikes causing civilian harm divided by 代表了军事行动对平民造成伤害的相对风险。随着时间的推移监测这一比率可以更好地了
解平民的相对风险是如何变化的，从而有可能针对新出现的和令人不安的趋势及早采取有重
the total number of strikes. This measure represents the relative risk of civilian harm from 点的干预措施。
military operations. Monitoring the rate over time allows for a better understanding of how
the relative risk to civilians is changing, enabling the possibility of early focused interventions
in response to emerging and troubling trends.
To support this process, CNA developed metrics to monitor the level of risk to civilians from 为了支持这一进程，CNA制定了指标来监测军事行动对平民的风险水平以及这种风险如何
military operations and how this risk changed over time. Part of our support to ISAF was 随着时间的推移而变化。我们对安援部队的部分支助是利用安援部队提供的数据每月监测
趋势，这使我们有机会实时应对平民面临的风险。例如2011年1月我们注意到几类行动的
monitoring trends each month with ISAF-provided data, which enabled an opportunity to
平民伤害率呈恶化趋势。在提醒安援部队注意这些趋势及其背后的因素之后，总部迅速作
respond to risks to civilians in real time. For example, in January 2011, we noticed worsening 出行动上的改变，以解决这些趋势的原因。结果，平民伤害的比率和数量又下降了。 22这
trends in the civilian harm rates for several types of operations. After alerting ISAF to these 种监测和应对新兴趋势的最佳做法后来被写入美国平民伤害国家政策。 23
trends and the factors behind them, the headquarters rapidly made operational changes that
addressed the causes of those trends. As a result, the rates and numbers of civilian harm went
back down.22 This best practice of monitoring and responding to emerging trends was later
written into US national policy for civilian harm.23
22Larry Lewis and Diane Vavrichek, Rethinking the Drone War, (Quantico, VA: Marine Corps University Press, 22Larry Lewis和Diane Vavrichek，重新思考无人机战争，（Quantico，VA：海军陆战队大学出版社，201
2016). 6）。
23 The White House, July 1, 2016, Executive Order--United States Policy on Pre- and Post-Strike Measures to Address 23白宫，2016年7月1日，行政命令-美国关于打击前措施的政策，以解决涉及使用武力的美国行动中的平民伤
Civilian Casualties in U.S. Operations Involving the Use of Force. The commitment to monitor and address trends is 亡问题。监测和处理趋势的承诺载于行政命令第4节。我们注意到，民警行动的速度取决于许多因素，例如行
contained in Section 4 of the Executive Order. We note that the rate of CIVCAS is dependent on many factors, such 动环境、故意危及平民的敌对战术和行动类型（例如空袭、炮火、地面行动），因此，如果一个行动的速度随
as operating environment, adversary tactics that purposely endanger civilians, and type of operation (e.g., air 着时间的推移而增加，或者一个行动的平均速度高于另一个行动的平均速度，就不一定是对军事力量的不良反
strikes, artillery fire, ground operation), so it is not necessarily a poor reflection on a military force if the rate 映。但至少，军事力量应该意识到并理解这些信息。
increases over time or if the average rate for one operation is higher than that for another operation. But at the
very least, a military force should be aware of and understand this information.
10 10
One example of identifying and responding to trends was our work on identifying and 确定和应对趋势的一个例子是，我们开展了工作，确定和建议采取步骤，以解决空袭期间
recommending steps to address specific risks of civilian harm during air strikes—a particular 平民受到伤害的具体风险--这是一个特别令人关切的领域，因为与其他类型的部队（例如
area of concern because air strikes tend to cause the most civilian harm per incident compared 小武器射击、炮火）相比，空袭往往造成的每个事件对平民的伤害最大。我们在这里讨论
我们确定的平民面临的三个风险领域以及在阿富汗行动过程中做出的调整：人员聚集，自卫
to other types of force (e.g., small arms fire, artillery fire). We discuss here three areas of risk 考虑和双重打击。
to civilians we identified and the adaptations that were made over the course of operations in
Afghanistan: gatherings of people, self-defense considerations, and double taps.
In Afghanistan, commander’s guidance (in the form of tactical directives) emphasized the need 在阿富汗，指挥官的指导（以战术指令的形式）强调需要减少平民伤害，包括根据未得到
to reduce civilian harm, including mitigating harm in light of the risk of unobserved civilians 服务的平民在建筑物中的风险减轻伤害。 2009年，该指南写道："我希望各级领导人根据这
being in buildings. In 2009, that guidance read: “I expect leaders at all levels to scrutinize and 一指南，仔细审查和限制对住宅小区和其他可能造成平民伤亡的地点使用武力，如近距离空
中支援。"24在我们对2009年至2010年在这一指导下发生的平民伤害事件的分析中，出现
limit the use of force like close air support against residential compounds and other locations 了一种模式：尽管部队对大院进行了更多的空袭，但几次空袭仍然针对不在建筑物内的人员
likely to produce civilian casualties in accordance with this guidance.”24 In our analysis of 聚集，指挥官的意图并没有始终如一地适用于这些袭击。随后的指导使这个案例更加清晰
civilian harm incidents in 2009 to 2010 that fell under that guidance, a pattern emerged: ，结果，空袭造成的平民伤害减少了，因为部队适应了减轻这种风险。
although forces exercised greater care in air strikes of compounds, several air strikes still
targeted gatherings of people not inside structures, and the commander’s intent was not being
applied consistently to those strikes. Subsequent guidance made this case clearer, and as a
result, civilian harm from air strikes were reduced because forces adapted to mitigate this risk.
Another common risk contributing to civilian harm occurred when strikes were made in self- 造成平民伤害的另一个常见风险是，在军队遭到炮火和要求空中支援的情况下，以自卫的
defense when military forces were under fire and calling for air support. Analysis of these cases 方式进行罢工。对这些案件的分析显示出一种模式：虽然罢工是由于紧急的自卫情况而开
revealed a pattern: although strikes began because of an urgent self-defense situation, attacks 始的，但在自卫情况不再存在之后，攻击有时仍在继续，尽管这些攻击并不是根据关于自卫
以外案件的交战规则首先批准的，因为这些案件需要更仔细地考虑附带损害问题。提供了
sometimes continued after the self-defense situation no longer existed, even though those 指导意见，指出自卫行动应设法只解决自卫情况，不应在未经事先批准的情况下转向攻击已
attacks were not first approved under the rules of engagement governing cases other than self- 查明的战斗人员的任务。结果，空袭对平民的伤害减少了。
defense, which require more careful consideration of collateral damage concerns. Guidance
was provided stating that self-defense engagements should seek to address only the self-
defense situation and should not shift to an offensive mission against identified combatants
without prior approval. As a result, civilian harm from air strikes were reduced.
civilian harm were also caused at times as part of the practice of “double taps.”. Some news 作为"双水龙头"做法的一部分，有时也会造成平民伤害。”. 一些新闻报道称，空袭将首先针
reports alleged that air strikes would first target combatants but then subsequently target 对战斗人员，然后针对急于援助幸存者的平民第一反应者。实际上，第二次空袭是根据对
civilian first responders rushing to aid the survivors. In actuality, the second air strikes were 第一次罢工没有执行任务的评估进行的，第一反应者由于与随后的罢工有关的附带损害估计
不足而受到伤害。对这些后续袭击提供更多指导，有时还要提供更多的监视能力，以帮助
based on assessments of the first strike not being mission effective, and first responders were 估计附带损害，从而减轻了这些情况下对平民的风险。
harmed because of an inadequate collateral damage estimate associated with the subsequent
strike. Additional guidance on these follow-on attacks, accompanied at times by additional
surveillance capabilities to aid collateral damage estimates, mitigated the risk to civilians in
those cases.
24NATO/ISAF, Tactical Directive, July 6, 2009. 24北约/安援部队，战术指令，2009年7月6日。 https://www.nato.int/isaf/

https://www.nato.int/isaf/docu/official_texts/Tactical_Directive_090706.pdf. docu/official_texts/Tactical_Directive_090706.pdf...
11 11
Other assessments 其他评估
Soon after we completed our work in Afghanistan supporting operations in real time, we 在我们完成在阿富汗支持实时行动的工作后不久，我们就平民伤害进行了两次评估。第一
conducted two assessments on civilian harm. The first examined specific risks to civilians from 个研究了使用无人机对平民的具体风险，第二个确定了美国行动减轻平民伤害的总体经验
the use of drones, and the second identified overarching lessons from US operations for 教训。 25这些经验教训为CNA起草2016年关于平民伤害的行政命令的两个部分的工作提供
了依据。 26
mitigating civilian harm.25 These lessons informed CNA’s work in drafting two sections of the
2016 Executive Order on civilian harm.26
CNA also led the analysis of operational data for several additional assessments, including the CNA还领导了对业务数据的分析，以进行一些额外的评估，包括2018年联合工作人员平民
2018 Joint Staff civilian harm Review and more recent assessments examining US- and 伤害审查和最近的评估，审查美国和联盟在伊拉克摩苏尔、叙利亚拉卡、阿富汗和索马里造
coalition-caused civilian harm in Mosul, Iraq; Raqqa, Syria; Afghanistan; and Somalia.27 In 成的平民伤害。 27此外我们审查了沙特领导的也门联盟在也门造成的平民伤害。 28我们还
评估了特定人群或群体的风险，例如儿童和医疗保健工作者。 29我们共分析了约2,000起现
addition, we examined civilian harm in Yemen caused by the Saudi-led coalition in Yemen.28 实世界的平民伤害事件。
We have also assessed risks to specific populations or groups, such as children and health care
workers.29 Collectively, we have analyzed about 2,000 real-world civilian harm incidents.
Root cause determinations: Why civilian harm 根本原因确定：为什么会发生平民伤害

happens
As we consider the concept of using AI to mitigate civilian harm, a key limitation is that AI is a 当我们考虑使用人工智能来减轻平民伤害的概念时，一个关键的限制是人工智能是解决特
tool for solving specific, well-defined problems. So, the problem of civilian harm must be well 定的，明确定义的问题的工具。因此，要有效地应用人工智能，必须对平民伤害问题进行
characterized to apply AI effectively. This limitation is not unique to AI and also applies to 良好的表征。这一限制并不是人工智能所独有的，也适用于军队：我们观察到，只有当军
队理解为什么首先发生平民伤害时，军队才能有效地减轻平民伤害。例如，在2006和2008
militaries: we have observed that military forces are effective in mitigating civilian harm only 之间的阿富汗，美国军方和国际部队试图采取措施减少平民伤害，但平民伤害事件的数量继
if the force understands why civilian harm occurs in the first place. For example, in Afghanistan 续攀升。虽然尝试了缓解措施，但效果不佳。我们对平民伤害的分析解释了为什么这些军
between 2006 and 2008, the US military and international forces attempted to institute 事措施不起作用：它们基于对平民伤害如何发生的错误假设。
measures to reduce civilian harm, but the number of civilian harm incidents continued to climb.
Although mitigation steps were attempted, they were not effective. Our analysis of civilian
harm explained why these military measures were not working: they were based on faulty
assumptions about how civilian harm happens.
25Larry Lewis, Drone Strikes: Civilian Casualty Considerations, Joint Staff, June 18, 2013; Lewis, Reducing and 25Larry Lewis《无人机打击平民伤亡考虑》联合参谋部2013年6月18日刘易斯《减少和减轻平民伤亡》
Mitigating Civilian Casualties. 。
26 The White House, July 1, 2016, Executive Order--United States Policy on Pre- and Post-Strike Measures. 26白宫，2016年7月1日，行政命令--美国关于罢工前和罢工后措施的政策。
27Department of Defense, Joint Staff CIVCAS Review, Apr. 17, 2018, https://www.justsecurity.org/wp- 27国防部，联合参谋CIVCAS审查，Apr。 2018年17月https://www.justsecurity.org/wpcontent/u
content/uploads/2019/02/Civ-Cas-Study-Redacted-just-security.pdf. ploads/2019/02/Civ-Cas-Study-Redacted-just-security.pdf...
28Larry Lewis, Learning from Yemen, CNA, IRM-2019-U-019749-Final, May 2019, 28Larry Lewis，向也门学习，CNA，IRM-2019-U-019749-决赛，2019年5月
https://www.cna.org/CNA_files/PDF/IRM-2019-U-019749-Final.pdf. ，https://www.cna.org/CNA_files/PDF/IRM-2019-U-019749-Final.pdf...
29For example, see Larry Lewis, Protecting Medical Care in Conflict: A Solvable Problem, CNA, Oct. 30, 2019, 29例如见Larry Lewis,Protecting Medical Care in Conflict:A Solvable Problem,CNA,Oct. 2019年30月https://hu
https://humanrightscommission.house.gov/sites/humanrightscommission.house.gov/files/documents/Protectin manrightscommission.house.gov/sites/humanrightscommission.house.gov/files/documents/Protectin g%20M
g%20Medical%20Care%20in%20Conflict%20-%20Lewis.pdf. edical%20care%20in%20conflict%20-%20lewis.pdf。
12 12
For example, in Afghanistan, the US military was operating under the assumption that when 例如，在阿富汗，美国军方在假设发生平民伤害时，军方正在从事有效的军事目标，该地
civilian harm occurred, the military was engaging a valid military target and civilians in the 区的平民在参与中受到伤害。我们称之为"附带损害"机制。但对现实世界事件的分析表明
area were harmed in the engagement. We call this a “collateral damage” mechanism. But ，大约有一半的时间，另一种机制正在发挥作用：军方确定了它认为是有效的军事目标并使
其相信，但目标实际上是平民。这种"错误识别"问题与"附带损害"问题有着根本的不同，需
analysis of real-world incidents showed that about half the time, another mechanism was in 要不同的解决方案。
play: the military identified what it believed was a valid military target and engaged it in that
belief, but the target was, in fact, civilians. This “misidentification” problem is fundamentally
different from the “collateral damage” problem and demands different kinds of solutions.
To better define the specific problems that AI applications should seek to solve, we performed 为了更好地定义人工智能应用程序应该寻求解决的具体问题，我们对我们研究的数千起事
a meta-analysis of the thousands of incidents we have examined, defining the specific 件进行了元分析，确定了过去15年导致平民伤害的具体机制，包括阿富汗、伊拉克、叙利
mechanisms that led to civilian harm over the last decade and a half, including in Afghanistan, 亚、索马里和也门。基于这一荟萃分析，我们开发了一个图形，如图1所示，说明了可能造
成平民伤害的不同机制。
Iraq, Syria, Somalia, and Yemen. Based on this meta-analysis, we developed a graphic, shown
in Figure 1, illustrating the different mechanisms that can cause civilian harm.
13 13
Figure 1. Mechanisms for civilian harm 图1。平民伤害机制
Source: CNA. Source: CNA.
14 14
As we stated above, there are two main causes of civilian harm: misidentification and collateral 如上所述，造成平民伤害的主要原因有两个：错误识别和附带损害。 30每个主要原因都有
damage.30 Each main cause has several variations. For example, misidentifications can be 几个变化。例如，错误识别可能是由于基于外表、行为或其他信息将平民误认为是威胁，
either from a misperception of civilians as being a threat based on appearance, behavior, or 也可能是由于错误地将有关有效军事目标的信息错误地归因于平民，从而认为他们实际上是
目标。
other information, or from a misassociation, where information about a valid military target is
wrongly ascribed to civilians, resulting in the belief that they are, in fact, the target.
Civilian harm through collateral damage can also be from a variety of causes. First, such harm 通过附带损害造成的平民伤害也可能来自各种原因。首先，这种损害可能已纳入相称性分
may have been factored into the proportionality analysis. However, most civilian harm we have 析。然而，我们所观察到的大多数平民伤害都来自意料之外的因素。一种是意料之外的存
observed was from factors that were unanticipated. One kind is unanticipated presence, where 在，在接触之后才观察到平民存在。这可以通过几种方式发生。对于预先计划的行动，平
民可能在场，但在附带损害估计中没有出现。例如，平民可能存在于建筑物或车辆中，但
civilian presence is not observed until after the engagement. This can happen in a few ways. 没有被军队或传感器观察到。或者，在进行附带损害估计时，可能没有平民在场，但平民
For preplanned operations, civilians may have been present but were missed in the collateral 后来进入该地区，没有人注意到。对于动态行动来说，如果交战过程被压缩，而且往往缺
damage estimate. For example, civilians may be present in buildings or vehicles but not 乏有意交战的规划和正式的目标批准程序，也有可能错过平民的存在。此外，无论交战是
observed by military forces or sensors. Alternately, no civilians may have been present at the 预先计划的还是动态的，当平民在交战期间进入目标地区时，都会发生平民伤害。我们称
time the collateral damage estimate was performed, but civilians moved into the area later, 之为短暂的平民存在。
unobserved. For dynamic operations, where the engagement process is compressed and often
lacks the planning and formalized target approval process of deliberate engagements, it is also
possible to miss the presence of civilians. Also, whether the engagement is preplanned or
dynamic, civilian harm can occur when civilians move into the target area around the time of
the engagement. We call this transient civilian presence.
The final type of collateral damage is due to unanticipated effects, of which we see three kinds. 最终类型的附带损害是由于意料之外的影响，其中我们看到三种。首先是武器装备和武器
The first is weaponeering and weapon performance issues that lead to the target being missed, 性能问题，导致目标被遗漏，导致受影响地点的平民伤害。另一个问题是武器制造和武器
性能问题，这些问题影响到预定区域，但也对平民存在的邻近地区或结构产生意想不到的影
resulting in civilian harm at the affected location. Another is weaponeering and weapon
响。最后，二次爆炸可能导致武器本身影响范围之外的平民受到伤害。
performance issues that affect the intended area but also have unanticipated effects on
adjacent areas or structures where civilians are present. Finally, secondary explosions can lead
to harm to civilians outside the range of effects of the weapon itself.
A civilian protection life cycle for mitigating 减少风险的平民保护生命周期

risks
Now that we have defined specific mechanisms for how civilian harm can occur, we can start 既然我们已经确定了如何发生平民伤害的具体机制，我们就可以开始考虑减轻伤害风险的
to consider possible steps to mitigate the risk of harm. In our work, we have found that 可能步骤。在我们的工作中，我们发现，减轻平民的风险应该考虑的远远超过"触发拉动"
接触决定。相反，我们发现，通过观察模式并采取措施减轻这些模式的适应性方法，减轻
mitigating risk to civilians should consider far more than the “trigger pull” engagement
平民伤害得到了加强。虽然个人
decisions. Rather, we find civilian harm mitigation is strengthened through an adaptive
approach to observing patterns and taking steps to mitigate them. Although individual
30 Here we are referring to direct harm to civilians that is accidental or inadvertent. Although in some cases 30在这里，我们指的是意外或无意对平民造成的直接伤害。虽然在某些情况下，平民伤害可以被认为是合
civilian harm can be accepted as lawful and proportional in an engagement of a valid military target, most cases 法的，并且在有效的军事目标的参与中是成比例的，但我们审查的大多数情况都没有在参与决定中预见到平
we have reviewed did not anticipate civilian harm in the engagement decision. 民伤害。
15 15
practices and policies can be beneficial for protecting civilians, the most impact is achieved 做法和政策可以有利于保护平民，最大的影响是通过我们称之为平民保护生命周期（CPLC
through a comprehensive approach we refer to as the civilian protection life cycle (CPLC). This ）的综合方法实现的。这一生命周期反映了在规划和使用军事力量的所有方面对平民保护
life cycle reflects attention to civilian protection at all points in the planning and use of military 的关注，并包括学习循环，以便军队能够适应和改进以克服挑战。图2所示的CPLC由以下
元素组成:
force and includes learning loops so that militaries can adapt and improve to overcome
challenges. The CPLC, illustrated in Figure 2, consists of the following elements:
• Mission and Mandate. This element involves designating and allocating needed • 任务和任务。这一要素涉及指定和分配必要的能力和权力，以便以从一开始就考
capabilities and authorities to conduct operations in ways that consider the protection 虑保护平民的方式开展行动。例如，可以对特派团进行塑造以减轻平民的风险，可
of civilians from the beginning. For example, the mission can be shaped to mitigate 以分配支持平民减轻伤害（CHM）要求的能力，并且可以调整权限和接战规则以减
轻平民的风险。
risks to civilians, capabilities to support civilian harm mitigation (CHM) requirements
can be allocated, and authorities and rules of engagement can be tailored to mitigate
risks to civilians.
• Planning. At strategic down to tactical levels, this element involves conducting • 规划。从战略到战术层面，这一要素包括进行规划和制定指挥指导，考虑到平民
planning and developing command guidance that factors in risks to civilians and 的风险，并包括帮助减轻风险的可行步骤和备选办法。例子包括制定行动以减轻对
includes feasible steps and alternatives to help mitigate them. Examples include 平民的具体风险，制定战术替代方案以避免既定的伤害模式，调整和调整指挥指导
以更好地解决伤害模式，进行强调缓解措施的任务排练，在适当时为CHM分配能力
shaping operations to mitigate specific risks to civilians, developing tactical ，以及确定平民的生活模式。
alternatives to avoid established patterns of harm, tailoring and adapting command
guidance to better address patterns of harm, conducting mission rehearsals that
emphasize mitigation measures, allocating capabilities for CHM when appropriate,
and determining pattern of life for civilians.
• Operational Execution. This element involves performing targeting processes that • 操作执行。这一要素涉及执行目标确定程序，促进准确识别和提供致命影响，同
promote accurate identification and delivery of lethal effects while seeking ways to 时寻求尽量减少平民伤害和回响影响的方法。例子包括行使战术耐心，故意武器化
以减轻平民风险，协调细节以确保共同目标并避免误联系，以及筛选短暂的平民存
minimize civilian harm and reverberating effects. Examples include exercising tactical
在。
patience, deliberate weaponeering to mitigate civilian risks, coordinating details to
ensure a common target and avoid misassociation, and screening for transient civilian
presence.
• Assessment. This element involves considering all available information, with • 评估。这一要素涉及考虑所有现有的信息，包括内部和外部来源，以确定对使用
internal and external sources, to determine the best estimate of civilian harm caused 武力造成的平民伤害的最佳估计。例子包括评估战斗损害以包括对平民的影响和监
by the use of force. Examples include assessing battle damage to include effects on 测潜在的平民伤害事件。
civilians and monitoring for potential civilian harm incidents.
• Response. This element involves working to mitigate the tragic consequences of • 回应。这一要素涉及努力减轻平民对受影响个人和人口的伤害的悲惨后果，包括
civilian harm to affected individuals and populations, including the provision of urgent 提供紧急医疗服务，向受害者作出补偿，并在适当时予以承认和道歉。
medical care, making amends to victims, and acknowledgment and apology when
appropriate.
• Learning and Adapting. This element involves using assessments, including patterns • 学习和适应。这一要素涉及使用评估，包括伤害模式和趋势数据，以确定业务改
of harm and trend data, to identify operational refinements to better protect civilians. 进，以更好地保护平民。这些评估还确定了有助于应对观察到的挑战的体制要求
These assessments also identify institutional requirements that can help address 。示例包括提供增强当前缓解的见解
observed challenges. Examples include giving insights to enhance current mitigation
16 16
efforts, sharing lessons and data for effective learning, and informing needed 努力，分享经验教训和数据以有效学习，并告知所需的体制改进。
institutional improvements.
• Institutional Capacity. This element involves designing the force to reduce risks to • 机构能力。这一要素涉及设计部队以减少对平民的危险，并应对整个军事机构所
civilians and addressing observed challenges and requirements across the military 观察到的挑战和要求（例如，理论、训练、物资解决方案），以加强随着时间的推
institution (e.g., doctrine, training, materiel solutions) to strengthen the ability to 移减轻对平民的伤害的能力。
mitigate harm to civilians over time.
Figure 2. Civilian protection life cycle 图2。平民保护生命周期
The CPLC also includes two learning loops: operational learning, where assessments of causes CPLC还包括两个学习循环：业务学习，对原因和趋势的评估直接为在正在进行的业务范围
and trends directly inform the improvement of operational practices and policies within the 内改进业务做法和政策提供信息；机构学习，对挑战和要求的评估为理论、政策、组织、培
context of an ongoing operation, and institutional learning, where assessments of challenges 训和领导以及设备和设施等方面的必要变化提供信息。
and requirements inform needed changes to, for example, doctrine, policy, organization,
training, and leadership, together with equipment and facilities.
Examples of practical steps for mitigating 减轻平民伤害的实际步骤的例子

civilian harm
Militaries can take various practical steps to mitigate civilian harm more effectively. Some are 军队可以采取各种实际步骤，更有效地减轻平民伤害。有些是普遍的，而另一些则是特定
universal, whereas others are particular to a specific mechanism of civilian harm. For example, 的平民伤害机制。例如，三种最佳实践可能适用于所有操作:
three best practices can potentially be applied to all operations:
17 17
• Tactical patience. Tactical patience involves taking time to verify positive • 战术耐心。战术耐心包括在情况允许的情况下，在攻击之前花时间验证对操作环
identification and understanding of the operating environment before attacking, when 境的积极识别和理解。
the situation allows.
• Tactical alternatives. Tactical alternatives involves considering different options for • 战术替代方案。战术替代方案涉及考虑不同的选择，以实现预期的效果，鉴于潜
achieving desired effects in view of potential second-order effects. 在的二阶效应。
• Shaping. Shaping involves planning and maneuvering forces to reduce the likelihood • 塑造。塑造涉及规划和机动部队，以减少可能需要在平民在场的情况下使用大量
of a situation in which significant force in the presence of civilians might be required. 部队的可能性。
We discuss each of these best practices below. 我们将在下面讨论这些最佳实践中的每一个。
Tactical patience can be applied in a range of environments and missions when it is consistent 战术耐心可以在一系列环境和任务中应用，当它与满足任务目标和遵循自卫考虑一致时。
with meeting mission objectives and following self-defense considerations. Some situations 有些情况需要立即使用武力：例如，一个高度优先的目标有失去的危险，或者军队正处于针
对眼前威胁的自卫情况。在其他情况下，有时间进一步考虑，部队可以使用战术耐心作为
require an immediate use of force: for example, a high-priority target is in danger of being lost
预防措施。战术耐心的真实世界示例包括以下内容:
or military forces are in a self-defense situation against an immediate threat. In other
situations, there is time for further consideration, and forces can use tactical patience as a
precautionary measure. Real-world examples of tactical patience include the following:
• Two pen flares. A soldier at a checkpoint aims a warning pen flare at a car that has • 两个笔照明弹。一名士兵在检查站将警告笔对准一辆没有注意到先前口头和视觉
not heeded earlier verbal and visual warnings. The car continues toward the soldier. 警告的汽车。汽车继续向士兵走去. 注意到朦胧的天气可能会阻碍能见度，士兵决
Noting hazy weather that could hinder visibility, the soldier decides he has time to fire 定他有时间发射另一个笔照明弹，而不是立即向汽车射击作为威胁。汽车司机看到
第二支笔闪动并停止，避免了平民的伤害。
another pen flare instead of resorting immediately to firing at the car as a threat. The
car driver sees the second pen flare and stops, averting civilian harm.
• Children in the road. An attack helicopter observes two individuals digging in a road • 路上的孩子。一架攻击直升机观察到两个人在一条道路上挖掘，并认为他们可能
and believes they could be laying an improvised explosive device along a road that 沿着军队经常行驶的道路铺设简易爆炸装置。由于没有立即攻击的迫切需要，直升
military forces often travel. Because there is no pressing need to attack immediately, 机重新定位以从不同的有利位置观看场景。从这个不同的角度来看，这两个人显然
是在路上挖掘的孩子。直升机不开火。
the helicopter repositions to view the scene from a different vantage point. Viewed
from this different perspective, the two individuals are clearly children digging in the
road. The helicopter does not fire.
Likewise, forces can plan for and employ tactical alternatives to mitigate risks to civilians from 同样，部队可以计划和采用战术替代方案，以减轻其行动对平民的风险。为了响应指挥部
their operations. In response to command emphasis on being effective and yet sparing civilians 强调在阿富汗有效和不放过平民的要求，那里的部队积极寻求解决办法，减少其任务中给平
in Afghanistan, forces there actively sought to find solutions that presented fewer risks to 民带来的风险，同时保持任务的成功和部队的安全。例如，一个单位将其方法从进行突袭
以抓住敌军转变为与当地部队合作进行普查行动，加深对当地情况的了解，同时也剔除了躲
civilians in their missions while preserving the success of the mission and the safety of the 藏在人口中的战斗人员。另一个单位从在火力下使用空袭进行近距离空中支援转变为使用
force. For example, one unit switched its approach from conducting raids to catch enemy forces 预先定位的狙击手，以减少空袭建筑物时可能发生的平民伤害的风险。
to conducting census operations in partnership with local forces, building better
understanding of the local situation while also culling out combatants hiding within the
population. Another unit shifted from using air strikes for close air support when under fire to
using pre-positioned snipers to reduce the risk of civilian harm that could occur during air
strikes on buildings.
In shaping, forces consider civilian harm risks in planning, equipment to employ, and the 在塑造方面，部队在规划、使用设备以及部队的部署和行动中考虑平民伤害风险，以便在
placement and movement of forces to reduce those risks when possible. This practice can be 可能的情况下减少这些风险。这种做法可以是
18 18
as easy as reconsidering the placement of a checkpoint, for example moving a checkpoint 就像重新考虑检查点的位置一样简单，例如将位于曲线周围的检查点移动到具有更多可见
positioned around a curve to a location with more visibility to allow more reaction time and 性的位置，以允许更多反应时间并启用战术耐心。这种做法还可能涉及制定整个行动概念
enable tactical patience. This practice can also involve shaping the entire concept of operation ，以更好地解决平民面临的风险，例如携带更适合威胁和环境的弹药或监视资产，或允许战
斗人员离开城市地区，以减少人口稠密地区的战斗强度，然后在其他地点追捕。
to better address risks to civilians, such as bringing in munitions or surveillance assets better
suited for the threat and environment or allowing fighters to leave an urban area to reduce the
intensity of fighting in highly populated areas and pursuing them later in other locations.
Some mitigation steps are particular to specific mechanisms. For example, for checkpoint 一些缓解措施是针对特定机制的。例如，在检查站行动中，平民受到伤害的主要原因是基
operations, civilian harm occurred primarily because of misidentifications based on behavior: 于行为的错误识别：平民的行为可能具有威胁性，或与预期威胁的行为相一致。车辆在接
civilians could appear to behave in threatening ways or in ways consistent with the behavior 近军队时没有响应减速的警告，这通常是因为典型的警告方法（例如，挥手和警告）在尘土
飞扬，能见度有限的情况下不能非常有效地引起驾驶员的注意。人为因素的失败也有所贡
of the anticipated threat. Vehicles did not respond to warnings to slow down as they 献，例如军队使用激光炫目器来警告司机，但颜色（绿色）被平民解释为继续前进的信号，
approached military forces, often because typical warning methods (e.g., waving and warning) 因此当他们没有停下来时就会被射击。
were not very effective in catching the attention of drivers in dusty, limited-visibility
conditions. Human factor failures also contributed, such as when military forces used laser
dazzlers intended to warn drivers but the color (green) was interpreted by civilians as a signal
to proceed, so they were fired upon when they did not stop.
Finding ways to more effectively communicate with civilian drivers reduces the problem of 找到更有效地与民用驾驶员沟通的方法，减少了基于行为的错误识别问题。这种沟通可以
misidentification based on behavior. This communication could include dedicated signs, 包括专门的标志，道路上的隆隆声条以提醒司机，以及与当地居民就检查点的程序和期望进
rumble strips on the road to alert drivers, and preemptive communication with local 行先发制人的沟通。了解这一机制也可以避免采取非最佳的缓解措施：例如，阿富汗军队
停止了警告射击，因为这些射击导致平民伤害的少数情况，但分析表明，取消这一步骤剥夺
populations regarding the procedures and expectations for checkpoints. Understanding this 了部队警告平民的唯一手段之一，并导致平民伤害净增加。
mechanism could also avoid nonoptimal mitigation steps: for example, military forces in
Afghanistan discontinued warning shots because of a few instances when those shots led to
civilian harm, but analysis showed that removing this step deprived forces of one of their only
means of warning civilians and led to a net increase in civilian harm.
Similarly, analysis of artillery fire showed how existing doctrine is geared toward high-end 同样，对炮火的分析表明，现有理论是如何针对高端冲突的，其中火炮放置的考虑集中在
conflict, where considerations of placement of artillery focus on the risk of friendly fire. Thus, 友好火力的风险上。因此，炮弹首先向友军相反的方向超过目标，然后将火力调整回预定
artillery rounds are fired to first overshoot the target in the direction opposite friendly forces 目标。虽然这种做法在没有平民的战场上是有意义的，但在城市环境中，它会给平民带来
重大风险。这种做法通过附带损害造成平民的伤害，这是由于火灾区内平民的存在没有被
and to then adjust fires back towards the intended target. Although this approach makes sense 注意到，这种情况因为减轻友军火力而制定的战术而加剧，但不考虑平民的存在。在附近
in a battlefield absent of civilians, in an urban setting, it can introduce significant risk to 没有友军的情况下，更好的方法是首先寻找潜在平民存在的迹象，然后向相反的方向发射初
civilians. This approach represents civilian harm through collateral damage due to the 始火力。
unnoticed presence of civilians in the area of fire, which is exacerbated by the tactic developed
to mitigate friendly fire but not to consider the presence of civilians. A better approach, in the
absence of friendly forces in proximity, is to first look for signs of potential civilian presence
and then place initial fires in the opposite direction.
19 19
Potential civilian harm mitigation measures AI的潜在平民伤害缓解措施
for AI
We have discussed the potential mechanisms whereby civilian harm occurs and provided a 我们讨论了发生平民伤害的潜在机制，并为如何采取缓解措施以减少对平民的风险提供了
framework for how mitigation steps can be introduced to reduce risks to civilians. But not all 一个框架。但并非所有这些步骤都适用于支持AI的应用程序。例如，考虑战术替代方案和
these steps are amenable to AI-enabled applications. For example, considering tactical 提供运营环境的塑造都是高度复杂的开放式任务，不适合AI可以执行的强大但狭隘的任务。
然而，战术耐心--如果存在某些风险因素或不满足某些要求，则暂停参与--可能是启用AI
alternatives and providing shaping of the operational environment are both highly complex, 的应用程序可以执行的功能。接下来的部分描述了我们认为是平民伤害问题的重要贡献者
open-ended tasks that are not well suited for the powerful but narrow nature of tasks that AI 的其他一些潜在缓解措施，这些措施的定义足够好，AI应用程序可能会成功。
can perform. However, tactical patience—pausing an engagement if certain risk factors are
present or if certain requirements are not met—is possibly a function that an AI-enabled
application could perform. The next sections describe some other potential mitigation
measures that we see as significant contributors to the problem of civilian harm that are
defined well enough that an AI application could potentially be successful.
Detecting transient civilian presence 检测瞬态平民存在

Transient individuals or vehicles moving into the engagement area undetected was one of the 短暂的个人或车辆在未被发现的情况下进入交战区域是最近美国行动中造成平民伤害的主
leading causes of civilian harm in recent US operations. In these cases, a person or vehicle 要原因之一。在这些情况下，一个人或车辆进入视野，在操作员注意到之前没有足够的时
moved into the field of view and there was insufficient time to abort the attack or steer the 间中止攻击或将武器转向更安全的区域。例如，在使用无人机时，操作员经常在接合前的
最后时刻放大，增加了瞬变的机会。这种现象在阿富汗早期的行动中也是一个问题，在那
weapon into a safer area by the time the operator noticed. For example, when using drones, 里部队制定了战术来试图减少这些事件：操作员将在交战决定之前将fullmotion视频分辨率
the operators often zoom in at the last moments before the engagement, increasing the chances 切换到更宽的视场，以更好地检测短暂的平民存在。
of transients. This phenomenon was also a problem in earlier operations in Afghanistan, where
forces developed tactics to try to reduce these occurrences: the operator would switch the full-
motion video resolution to a wider field of view before the engagement decision to better
detect transient civilian presence.
Recognizing protected symbols 识别受保护的符号

Hospitals and other humanitarian groups are protected from attack per international 根据国际人道主义法，医院和其他人道主义团体受到保护，不受攻击，但他们仍然经常受
humanitarian law, but they still are attacked all too often, sometimes because militaries fail to 到攻击，有时是因为军方不承认他们的身份。这些网站通常使用符号来直观地传达其受保
recognize them for what they are. These sites often use symbols to visually communicate their 护的性质。这些符号可以包括红十字会，红新月会和蓝十字（用于历史/文化遗址）。人类
操作员可能由于多种原因无法观察到这些符号：他们可能会错过他们对感知威胁的关注，或
protected nature. These symbols can include the red cross, the red crescent, and the blue cross 者他们可能正在通过红外瞄准吊舱进行观察，这使得
(for historical/cultural sites). Human operators can fail to observe these symbols for multiple
reasons: they may miss them in their focus on a perceived threat or they may be viewing
through an infrared targeting pod that makes it harder to clearly differentiate colored objects.
20 20
Developing a robust civilian pattern of life 发展稳健的平民生活模式
For deliberate, preplanned attacks, militaries often conduct a pattern of life assessment and a 对于蓄意的、预先计划的攻击，军队经常使用可用的图像进行生命评估和附带损害估计。
collateral damage estimate using available imagery. If the attack is not conducted until days, 如果攻击直到评估后几天，几周或更长时间才进行，地面条件可能会发生变化。此外，生
weeks, or longer after the assessment, the conditions on the ground can change. Also, pattern 活评估的模式本质上是有限的。例如，在评估的一段时间内，民事活动可能并不明显。此
外，这些评估倾向于考虑威胁地点，而不是平民或人道主义地点，倾向于对威胁作出决定，
of life assessments are inherently limited. For example, civilian activity may not be evident in 而不是将其与有关平民和人道主义实体的其他现有信息进行平衡。
the slice of time of the assessment. Also, these assessments tend to factor in threat locations
but not civilian or humanitarian locations, tending to sway determinations toward a threat
instead of balancing them with other available information regarding civilian and
humanitarian entities.
Improving collateral damage estimates 改善附带损害估计

For militaries that have a formal collateral damage estimation process, this process gives an 对于有正式附带损害估计过程的军队，这个过程给出了一个可能的平民伤亡的想法，可以
idea of the potential civilian toll that can be expected from a particular attack. But these 预期从一个特定的攻击。但这些估计可能会出现严重不足。一个不足之处是附带损害估计
estimates can suffer from severe shortfalls. One shortfall is that the collateral damage estimate 值表示时间上的离散快照。如果事情在估计时间和订婚时间之间发生变化怎么办？没有标
准的程序或工具来评估是否发生了可能影响预计平民死亡人数的变化。另一个不足之处是
represents a discrete snapshot in time. What if things change between the time of the estimate ，这些估计是基于一个标准的统计模型，而冲突地区平民的生活模式可能因稳定状态而受到
and the time of the engagement? There is no standard process or tool that evaluates whether 严重破坏。假设平民只是逃离冲突地区是不安全的。经证实的涉及躲藏在激烈战斗地区建
a change has occurred that may affect the projected civilian toll. Another shortfall is that these 筑物内的家庭的平民伤害事件表明，即使军方认为并非如此，平民也可能在场。值得注意
estimates are based on a standard statistical model, whereas the pattern of life of civilians in 的是，附带损害估计模型没有根据实际操作结果进行校准和完善。
conflict areas can be significantly disrupted from steady-state conditions. And it is not safe to
assume that civilians will simply flee conflict areas. Confirmed civilian harm incidents
involving families hiding in buildings in areas of heavy fighting show that civilians may be
present even when military forces believe otherwise. Notably, models for collateral damage
estimation are not calibrated and refined by actual operational results.
Detecting misassociations Detecting misassociations

One contributing factor to civilian harm is misasociation, where a surveillance platform follows 造成平民伤害的一个因素是误关联，即监视平台跟随正确识别的战斗人员，但随着时间的
a correctly identified combatant but over time the target is “swapped” with a civilian entity. 推移，目标与平民实体"交换"。军事监视没有注意到交换，导致平民误以为他们是军事目
The military surveillance does not notice the swap, leading to the engagement of civilians in 标。这种情况经常发生，因为随着时间的推移，没有对目标的性质进行严格的审查。是同
一辆车还是一群人? 他们还有相同的情报特征吗？另一个因素是对目标位置的混淆，例如，
the mistaken belief that they are a military target. This situation often occurs because there is 当正确识别的威胁位于一个位置时，该目标被传递给其他人，然后其他人在不同的位置看到
no critical scrutiny of the nature of the target over time. Is it the same vehicle or group of 另一个实体并将其误
people? Do they still have the same intelligence signature? Another factor is confusion over the
target location, for example when the correctly identified threat is at one location and that
target is handed off to someone else who then sees another entity at a different location and
mistakes them as the threat.
21 21
Deconflicting with critical infrastructure 与关键基础设施解除冲突
Although military strikes increasingly include consideration of immediate collateral damage to 虽然军事打击越来越多地包括考虑对民用结构的直接附带损害，但它们往往没有考虑对关
civilian structures, they often fail to consider the impact on critical infrastructure. Damage to 键基础设施的影响。对这种基础设施(例如电力、水)的破坏可能对平民人口产生广泛和持久
such infrastructure (e.g., electricity, water) can have widespread and lasting effects on civilian 的影响。理想情况下，军方会确定关键基础设施的位置，然后标记在这些位置或附近的潜
在攻击，但实际上并不会发生这种情况。
populations. Ideally militaries would identify locations of critical infrastructure and then flag
potential attacks at or near those locations, but this does not tend to happen in practice.
22 22
Leveraging AI Applications for 利用人工智能应用程序
Mitigating Civilian Harm 减轻平民伤害
Having introduced the problem of civilian harm and some specific challenges that increase risk 在介绍了平民伤害问题和增加平民风险的一些具体挑战之后，我们现在讨论如何利用人工
to civilians, we now discuss how AI and ML can be leveraged to mitigate civilian harm. Our 智能和ML来减轻平民伤害。我们的讨论有三个部分。我们首先确定每个CPLC的六个基本
discussion has three parts. We first identify specific actions in each of the six basic components 组成部分的具体行动，着眼于可能适用于AIor ML启用的方法的行动。其次，以类似的方式
，我们分析了AI/ML应用程序的空间，以确定那些最有希望解决CPLC关键要素的应用程序，
of the CPLC with an eye toward actions that are potentially amenable to AI- or ML-enabled 优先考虑直接解决平民伤害机制和我们在上一章中确定的相关系统挑战的应用程序。最后
approaches. Second, in analogous fashion, we parse the space of AI/ML applications to identify ，我们将这些努力结合起来，明确地将潜在的平民伤害缓解清单与一组特定的AI/ML方法和
those that are most promising for addressing key elements of the CPLC, prioritized by 技术联系起来。
applications that directly address mechanisms for civilian harm and associated systemic
challenges we identified in the previous chapter. Finally, we combine these efforts by explicitly
associating a list of potential civilian harm mitigations with a set of specific AI/ML methods
and technologies.
Connecting the CPLC with AI/ML applications 连接CPLC与AI/ML应用

Figure 3 shows the CPLC−AI/ML applications (CPLC-AI) matrix, which associates CPLC 图3显示了CPLC−AI/ML应用（CPLC-AI）矩阵，其将CPLC元件与特定的AI/ML应用相关联。
elements with specific AI/ML applications. The 26 rows of this matrix denote specific CPLC 该矩阵的26行表示特定的CPLC操作（与六个主要CPLC元素中的每一个相关联，以绿色突出
显示），而列表示33个基本类AI和ML应用程序的分类（如下所述）（组织成11个"顶级"域
actions (associated with each of the six main CPLC elements, highlighted in green), whereas
，以蓝色突出显示，如自主无人飞行系统（UASs）/无人水面舰艇（Usv）网络空间和计算
the columns represent a taxonomy (described below) of 33 basic classes of AI and ML 机视觉）。如果给定的AI/ML应用程序（例如，列索引A）对并且可以合理地利用以帮助支
applications (organized into 11 “top level” domains, highlighted in blue, such as autonomous 持给定的CPLC操作（例如，行索引C），则（C，A）矩阵条目包含符号"◼";否则矩阵元素
unmanned aerial systems (UASs)/unmanned surface vessels (USVs) cyberspace, and computer 为空。以绿色和蓝色突出显示的括号中的数字分别表示出现在相应行和列中的"◼"条目的数
vision). If a given AI/ML application (e.g., column index A) is of value to and can be reasonably 量。
leveraged to help support a given CPLC action (e.g., row index C), then the (C, A) matrix entry
contains the symbol “◼”; otherwise the matrix element is empty. The numbers in parentheses
highlighted in green and blue denote the number of “◼” entries that appear in the
corresponding row and column, respectively.
Neither the AI/ML applications taxonomy nor the veracity of matrix elements is definitive (the AI/ML应用分类法和矩阵元素的真实性都不是确定的（国际社会的人工智能研究人员还没有
international community of AI researchers has not yet reached a consensus on how AI ought 就人工智能应该如何定义达成共识，更不用说确定一个普遍商定的方法和应用分类法了）。
31尽管如此，矩阵可以说抓住了本文中编织在一起的两个主要"概念空间"的核心格式塔，
to even be defined, much less settled on a universally agreed-upon taxonomy of methods and
即可以采取的行动来帮助减少平民伤害和可以利用的AI/ML技术请注意，虽然AI/ML分类法
applications).31 Nonetheless, the matrix arguably captures the core gestalt of the two main 绝不是神圣不可侵犯的（我们坚持三
“concept spaces” being woven together in this paper, namely actions that can be taken to help
reduce civilian harm and AI/ML technologies that can be leveraged to facilitate their efficacy.
Note that although the AI/ML taxonomy is by no means sacrosanct (our adherence to three
31 Andrew Ilachinski, AI, Robots, and Swarms, CNA, DRM-2017-U-014796-Final, Jan. 2017. 31Andrew Ilachinski,AI,Robots,And Swarms,CNA,DRM-2017-U-014796-Final,Jan. 2017.
23 23
applications per “top level” domain is arbitrary and was chosen mostly for expedience and to 每个"顶级"域的应用程序是任意的，主要是为了权宜之计和节省空间而选择的），分类法
save space), the top level of the taxonomy is based partly on a recent survey of state-of-the-art 的顶级部分基于最近对最先进方法的调查32，部分基于人工智能和ML在国家安全委员会（
methods32 and partly on how AI and ML are characterized in the National Security Commission Nscai）的最终报告中的特征。 33矩阵元素本身的值完全由本文作者裁决。虽然图3中单个"
◼"条目的存在（或不存在）可能会受到质疑，但最重要的是这种基质（或这种基质的似是
on Artificial Intelligence’s (NSCAI’s) final report.33 The values of the matrix elements 而非可信的变体）揭示结合CPLC元素和现有AI/ML应用
themselves were adjudicated entirely by the authors of this paper. Although the presence (or
absence) of individual “◼” entries in Figure 3 may be questioned, what matters most is the
degree to which this matrix (or plausibly credible variants of this matrix) reveals the overall
connective tissue that binds elements of the CPLC and extant AI/ML applications.
32See Appendix E (“AI/ML Approaches, Methods, and Algorithms Taxonomy”) in Andrew Ilachinski, Artificial 32见附录E"AI/ML方法、方法和算法分类学"Andrew Ilachinski《人工智能新兴主题、问题和叙述》CNA D
Intelligence: Emerging Themes, Issues, and Narratives, CNA, DOP-2020-U-028073-Final, Oct. 2020. op-2020-U-028073-Final10月10日。 2020.
33NSCAI, Final Report, Mar. 2021, https://www.nscai.gov/wp-content/uploads/2021/03/Full-Report-Digital- 33NSCAI，最后报告，3月。 2021年，https://www.nscai.gov/wp-content/uploads/2021/03/Full-Report
1.pdf. -Digital1.pdf...
24 24
Figure 3. CPLC-AI matrix 图3。 CPLC-AI矩阵
To better understand the CPLC-AI matrix, the next sections take a closer look at the AI/ML 为了更好地理解CPLC-AI矩阵，接下来的部分将详细介绍AI/ML应用分类法，从分布在顶行
applications taxonomy, starting with the domains spread across the top row (highlighted in （以蓝色突出显示）的域开始。每个域都描绘了AI或ML应用已经存在的区域，并且可以在
blue). Each domain depicts an area for which AI or ML applications already exist, and can be 近期中期对CPLC产生的问题产生影响。
brought to bear on problems germane to the CPLC in the near- to mid-term.
Autonomous UASs/UAVs Autonomous UASs/UAVs

This domain includes autonomous UASs, USVs, and drones in general. Unmanned systems are 这个领域包括自主UASs、Usv和一般的无人机. 无人系统是最公开可见的人工智能技术之一
among the most publicly visible AI-enabled technologies, spanning the spectrum from Defense ，涵盖了防御领域
25 25
Advanced Research Projects Agency (DARPA)-level challenges and prototypes34 to deployed 高级研究计划局(DARPA)级的挑战和原型34向部署的军事系统35(例如用于识别战场上感兴
military systems35 (e.g., systems used to identify targets of interest on the battlefield)36 to 趣的目标的系统)36向无人驾驶汽车。 37自主机器人可用于帮助支持CPLC要素的无数可能
driverless cars.37 Among the myriad possible ways autonomous robots can be used to help 方式包括损害评估与其他技术协同38提供援助和必需品39以及作为更大的网络传感器群的
一部分可 40
support elements of the CPLC are damage assessment (in concert with other technologies),38
delivery of aid and essentials,39 and as parts of larger networked sensor swarms (that can be
used to vastly enhance situational awareness (SA) over wide areas).40
In early 2020, shortly after the COVID-19 pandemic took hold of the world, the US-based drone 2020年初，在COVID-19大流行席卷全球后不久，美国无人机制造商Draganfly与澳大利亚
maker Draganfly partnered with the Australian Department of Defense and University of South 国防部和南澳大利亚大学合作，部署了特殊的"大流行无人机"，可以从远处检测咳嗽、打喷
Australia to deploy special “pandemic drones” that can detect coughing, sneezing, respiratory 嚏、呼吸频率甚至发烧。 41
rate, and even fever from a distance.41
Two recent research programs based in part on robot and swarm technologies are (1) DARPA’s 最近的两个部分基于机器人和群体技术的研究计划是（1）DARPA的攻击性群体启用战术计
OFFensive Swarm‐Enabled Tactics program, which seeks to develop swarms of collaborative 划，该计划旨在开发成群的协作自治系统来监视作战区域，建筑物和物体，为城市环境中的
autonomous systems to surveil operational areas, buildings, and objects to provide real-time 部队提供实时可操作的情报，42和（2）一个称为Roborder项目的大规模自治监视系统，4
3由无人移动机器人组成
actionable intelligence to troops in urban environments,42 and (2) a large-scale autonomous
surveillance system called the Roborder project,43 which consists of unmanned mobile robots
34 “DARPA Subterranean Challenge Announces Systems Competition Teams for Final Event,” DARPA website, May 34"DARPA地下挑战赛宣布系统竞赛团队参加决赛"，DARPA网站，2021年5月3日，https://www.darpa.mil/ne
3, 2021, https://www.darpa.mil/news-events/2021-05-03. ws-events/2021-05-03...
35Dan Gettinger, The Drone Databook, The Center for the Study of the Drone at Bard College, Sept. 2019, 丹*葛廷格，无人机数据库，巴德学院无人机研究中心，9月9日。 2019年，https://dronecenter.bard
https://dronecenter.bard.edu/files/2019/10/CSD-Drone-Databook-Web.pdf. .edu/files/2019/10/CSD-Drone-Databook-Web.pdf...
36Arthur Holland Michel, Unarmed and Dangerous: The Lethal Applications of Non-Weaponized Drones, The Center 36Arthur Holland Michel，无武装和危险：非武器化无人机的致命应用，巴德学院无人机研究中心，Mar。 20
for the Study of the Drone at Bard College, Mar. 2020, https://dronecenter.bard.edu/files/2020/03/CSD- 20年，https://dronecenter.bard.edu/files/2020/03/CSDUnarmed-and-Dangerous-Web.pdf...
Unarmed-and-Dangerous-Web.pdf.
37 Lawrence D. Burns, Autonomy: The Quest to Build the Driverless Car—And How It Will Reshape Our World, (New 37Lawrence D.Burns,Autonomy:The Quest To Build The Driverless Car-And How It Will重塑我们的世界,(New
York: Ecco Press, 2018). York:Ecco Press,2018).
38Xiaoyu Zhu, Junwei Liang, and Alexander Hauptmann, “MSNet: A Multilevel Instance Segmentation Network for 朱晓宇、梁俊伟和亚历山大*豪普特曼，"MSNet：空中视频自然灾害损害评估的多级实例分割网络"，12月38
Natural Disaster Damage Assessment in Aerial Videos,” Dec. 30, 2020, arXiv:2006.16479 [cs.CV]. 日。 30，2020，arXiv：2006.16479[cs.CV]。
39A notable recent example is the US company Zipline’s lightweight fixed-wing drones delivering blood to 25 39最近一个值得注意的例子是美国公司Zipline的轻型固定翼无人机在2019年向卢旺达各地的25家医院和诊所
hospitals and clinics across Rwanda in 2019. Evan Ackerman and Michael Koziol, “In the Air with Zipline’s Medical 输送血液。 Evan Ackerman和Michael Koziol，"在空中与Zipline的医疗交付无人机，"IEEE频谱（Apr。 30，201
Delivery Drones,” IEEE Spectrum (Apr. 30, 2019), https://spectrum.ieee.org/in-the-air-with-ziplines-medical- 9），https://spectrum.ieee.org/in-the-air-with-ziplines-medicaldelivery-drones/particle-1...
delivery-drones/particle-1.
40Dan Popescu et al, “A Survey of Collaborative UAV–WSN Systems for Efficient Monitoring,” Sensors 19, no. 21 40Dan Popescu et al,"A Survey of Collaborative UAV–WSN Systems for Efficient Monitoring,"Sensors19,no.
(Oct. 28, 2019), https://www.mdpi.com/1424-8220/19/21/4690/htm. 21(Oct. 28，2019），https://www.mdpi.com/1424-8220/19/21/4690/htm...
41“Can a Pandemic Drone Help Stop the Spread of COVID-19?” Draganfly website, July 3, 2020, 41"大流行无人机可以帮助阻止COVID-19的传播吗？"Draganfly网站，2020年7月3日，https://draganf
ly.com/news/can-a-pandemic-drone-help-stop-the-spread-of-covid-19/。
https://draganfly.com/news/can-a-pandemic-drone-help-stop-the-spread-of-covid-19/.
42“OFFensive Swarm-Enabled Tactics (OFFSET),” DARPA website, https://www.darpa.mil/program/offensive- 42"进攻群-启用战术(偏移),"DARPA网站,https://www.darpa.mil/program/offensiveswarm-enabled-tactics
swarm-enabled-tactics. Technical details appear in Timothy H. Chung, “OFFensive Swarm-Enabled Tactics 。技术细节出现在Timothy H.Chung，"OFFensive Swarm-Enabled Tactics（OFFSET），"Mar。 2021年，ht
(OFFSET),” Mar. 2021, https://apps.dtic.mil/sti/pdfs/AD1125864.pdf. tps://apps.dtic.mil/sti/pdfs/AD1125864.pdf...
43“Roborder: Autonomous Swarm of Heterogeneous Robots for Border Surveillance,” Roborder website, accessed 43"Roborder：用于边境监视的异构机器人的自主群"，Roborder网站，2021年9月20日访问，https://roborde
Sept 20, 2021, https://roborder.eu/wp-content/uploads/2021/03/ROBORDER_General_v2.0.pdf. r.eu/wp-content/uploads/2021/03/ROBORDER_General_v2.0.pdf...
26 26
(including aerial, ground, water surface, and underwater vehicles) capable of functioning （包括空中、地面、水面和水下飞行器）能够作为独立系统或使用多模式自适应传感器部
either as standalone systems or in swarms that are deployed with multimodal adaptive 署的成群系统运行。试点案例包括早期识别和跟踪非法活动和通信，以及发现边界发生的"
sensors. Pilot cases include early identification and tracking of illegal activities and 事故"。 44最近的一份文件对微型无人机用于遥感的使用情况进行了全面调查。 45
communications, and detection of “accidents” that occur at borders.44 A recent paper provides
a comprehensive survey of the use of mini-UAVs for remote sensing.45
Cyberspace Cyberspace
In the broadest sense, this domain denotes a set of applications derived principally from (or 从最广泛的意义上说，这个域表示一组主要来自（或以其他方式涉及或直接利用）通过计
that otherwise involve or directly leverage) data that are communicated over computer 算机网络通信的数据的应用程序。 46平民伤害相关的应用包括情报的生成和传播，47众包
映射技术48（也涉及计算机视觉），49和预测情感分析50（它结合了社交网络51和自然语
networks.46 civilian harm-related applications include the generation and dissemination of
言处理）。全球事件、语言和语调数据库提供了大量的数据，描述了人类在社会层面上随
intelligence,47 crowdsourced mapping technologies48 (which also involve Computer Vision),49 时间的行为；它的目的是监视"......世界上几乎每个国家每个角落的广播、印刷和网络新闻，
and predictive sentiment analysis50 (which combines social networks51 and Natural Language 用100多种语言，并确定了每天每一秒推动我们全球社会的人、地点、组织、主题、来源、
Processing). The Global Database of Events, Language, and Tone provides a vast set of data that 情感、计数、引言、图像和事件。” 52
describes human behavior at the societal level over time; it is designed to monitor “…the
world's broadcast, print, and web news from nearly every corner of every country in over 100
languages and identifies the people, locations, organizations, themes, sources, emotions,
counts, quotes, images and events driving our global society every second of every day.”52
44 “Roborder: Aims & Objectives,” Roborder website, https://roborder.eu/the-project/aims-objectives/. 44"Roborder:Aims&Objectives,"Roborder网站,https://roborder.eu/the-project/aims-objectives/。

45Tian-Zhu Xiang, Gui-Song Xia, and Liangpei Zhang, “Mini-Unmanned Aerial Vehicle-Based Remote Sensing: 45项天祝、夏桂松和张良培，"基于微型无人飞行器的遥感：技术、应用和前景"，IEEE地球科学与遥感杂
Techniques, Applications, and Prospects,” IEEE Geoscience and Remote Sensing Magazine 7, no. 3 (Sept. 2019). 志7，第3期（9月）。 2019).
46 Note that our use of the term “cyberspace” deliberately focuses on its data communication–centric meaning and 46请注意我们对"网络空间"一词的使用有意侧重于其以数据通信为中心的含义而不是处理网络安全问题的更典
not the more typical focus of issues dealing with cybersecurity. B. Geluvaraj, P. Satwik, and T. Kumar, “The Future 型的重点。 B.Geluvaraj，P.Satwik和T.Kumar，"网络安全的未来：人工智能，机器学习和深度学习在网络空间中
of Cybersecurity: Major Role of Artificial Intelligence, Machine Learning, and Deep Learning in Cyberspace,” S. 的主要作用，"S.Smys，R.Bestak，J.Z.Chen和i.Kotuliak，eds，in International Conference on Computer Netwo
rks and
Smys, R. Bestak, J.Z. Chen, and I. Kotuliak, eds, in International Conference on Computer Networks and
Communication Technologies, Singapore: Springer, 2019. 通信技术，新加坡：Springer，2019。
47Daniel Ish, Jared Ettinger, and Christopher Ferris, Evaluating the Effectiveness of Artificial Intelligence Systems in 47Daniel Ish,Jared Ettinger,And Christopher Ferris,Evaluating The Effectiveness Of Artificial Intelligence Syste
Intelligence Analysis, RAND Corporation, RR-A464-1, 2021. ms in Intelligence Analysis,RAND Corporation,RR-A464-1,2021.
48 Andrew Ilachinski, Applications of Social Media to Military Operations: Overview and Assessment, CNA, DME- 48Andrew Ilachinski，社交媒体在军事行动中的应用：概述和评估，CNA，DME2013-U-005368-Final，
2013-U-005368-Final, July 2013. 2013年7月。
49Kotaro Hara and Jon E. Froehlich, “Characterizing and Visualizing the Physical World Accessibility at Scale Using 49Kotaro Hara和Jon E.Froehlich,"Characterizing and Visualizing The Physical World Accessibility at Scale Using
Crowdsourcing, Computer Vision, and Machine Learning,” ACM SIGACCESS Accessibility and Computing 113 (Oct. Crowdsourcing,Computer Vision,And Machine Learning,"ACM SIGACCESS Accessibility and Computing113(Oc
2015). t. 2015).
50Yousef Mourabit et al., “A New Sentiment Analysis System of Tweets Based on Machine Learning Approach,” 50Yousef Mourabit等人。，"基于机器学习方法的推文新情感分析系统"，国际科技研究杂志9，第12期（1
International Journal of Scientific and Technology Research 9, no. 12 (Dec. 2020). 2月12日）。 2020).
51Qiaoyu Tan, Ninghao Liu, and Xia Hu, “Deep Representation Learning for Social Network Analysis,” Apr. 18, 谭巧玉，刘宁浩，夏胡，"社会网络分析的深度表示学习"，4月。 18，2019，arXiv：1904.08547[cs.SI]。
2019, arXiv:1904.08547 [cs.SI].
52 GDELT homepage, accessed Sept. 20, 2021, https://www.gdeltproject.org/. 52GDELT主页，9月访问。 2021年https://www.gdeltproject.org/。
27 27
The military’s interest in using social media as a new form of intelligence (social media 军方对使用社交媒体作为一种新形式的情报（社交媒体情报或SOCMINT）53的兴趣，特别
intelligence or SOCMINT)53 in general, and crowdsourcing54 in particular, dates to DARPA’s 是众包54，可以追溯到DARPA的2009红气球竞赛（也称为网络挑战赛）。 55.本次竞赛旨
2009 Red Balloon Contest (also known as the Network Challenge).55 The contest was designed 在探讨如何利用互联网和社交网络来解决分布式、时间关键的地理定位问题。具体来说，
挑战是找到部署在美国大陆未公开地点的10红色天气气球。麻省理工学院的一个媒体实验
to probe how the internet and social networking may be used to solve a distributed, time- 室团队通过使用社交媒体在九小时内找到了所有这些。除了展示社交媒体在解决现实世界"
critical geolocation problem. Specifically, the challenge was to find 10 red weather balloons 空间搜索"问题方面的效用（回顾2009年社交媒体刚刚起步时更令人印象深刻），结果还表
that were deployed at undisclosed locations across the continental United States. An MIT Media 明，尽管努力提供有关气球位置的虚假信息，但通过众包"解决方案"是可能的。 56
Lab team located all of them within nine hours by using social media. Apart from
demonstrating the utility of social media in solving a real-world “spatial search” problem (all
the more impressive when recalling that in 2009 social media was only nascent), the result also
showed that a “solution” was possible via crowdsourcing despite efforts to provide false
information on the location of the balloons.56
A widely used system that demonstrates the power of crowdsourcing is Ushahidi (which 展示众包力量的广泛使用的系统是Ushahidi（在斯瓦希里语中意为"证词"或"证人"），这是
means “testimony” or “witness” in Swahili), a freely available open-source platform for data 一个免费提供的开源平台，用于数据收集，可视化和交互式映射。它于2007年被引入，作
collection, visualization, and interactive mapping. It was introduced in 2007 as a one-stop 为肯尼亚2007年有争议的总统选举后暴力事件的一站式数据来源。这个想法是使用暴力的
目击者报告（通过电子邮件或短信发送）和谷歌地图来可视化特定暴力事件发生地点的物理
source of data about violence in Kenya following Kenya's disputed 2007 presidential election. 分布。从那时起，Ushahidi一直在不断发展，并被用于跟踪各种区域事件的演变，包括201
The idea was to visualize the physical distribution of locations where specific violent events 0年1月海地地震后的救灾工作，2011年叙利亚革命和2015年尼泊尔地震。 57
occurred using eyewitness reports of violence (sent in by email or text message) and Google
Maps. Since then, Ushahidi has been in continual development and has been used to track the
evolution of a variety of regional events, including the disaster relief efforts following the
earthquake in Haiti in January 2010, the Syrian revolution in 2011, and the earthquake in Nepal
in 2015.57
As another illustrative example of the power of crowdsourcing, a recent paper in Armor: 作为众包力量的另一个说明性例子，最近的一篇论文装甲：安装机动杂志引用了2017年的
Mounted Maneuver Journal cited an event from 2017 at which 1st Brigade Combat Team, 4th 一个事件，1旅战斗队，4步兵师正在国家训练中心进行侦察。该旅成功地改变了其领导营
Infantry Division was conducting a reconnaissance-in-force at the National Training Center. 的方向，以避免在一个窒息点的伏击。
The brigade successfully redirected its lead battalion to avoid an ambush at a chokepoint based
53Bruce Forrester and Kees den Hollander, “The Role of Social Media in the Intelligence Cycle,” in Next-Generation 53Bruce Forrester和Kees den Hollander，"社交媒体在情报周期中的作用"，载于Next-Generation Analyst IV，
Analyst IV, Proceedings Vol. 9851, May 12, 2016. Proceedings Vol。 9851，2016年5月12日。
54Kathryn B. Laskey, “Crowdsourced Decision Support for Emergency Responders,” in 18th International 54Kathryn B.Laskey，"紧急响应者的众包决策支持"，在第18届国际指挥和控制研究与技术研讨会，弗
Command and Control Research and Technology Symposium, Alexandria, VA, June 19-21, 2013, 吉尼亚州亚历山大，2013年6月19-21日，https://apps.dtic.mil/sti/pdfs/ADA588344.pdf...
https://apps.dtic.mil/sti/pdfs/ADA588344.pdf.
55John C. Tang et al., “Reflecting on the DARPA Red Balloon Challenge,” Communications of the ACM 54, no. 4 (April 55John C.Tang等人。，"反思DARPA红气球挑战"，ACM54通讯，第4号（2011年4月），https://cacm.acm.org
2011), https://cacm.acm.org/magazines/2011/4/106587-reflecting-on-the-darpa-red-balloon-challenge/fulltext. /magazines/2011/4/106587-reflecting-on-the-darpa-red-balloon-challenge/fulltext...
56Alex Rutherford et al., “Impossible by Conventional Means: Ten Years on from the DARPA Red Balloon 56Alex Rutherford等人。，"传统方法不可能：十年后的DARPA红气球挑战，"8月8日。 13，2020，
Challenge,” Aug. 13, 2020, arXiv:2008.05940v1,. arXiv：2008.05940v1，。
57Sergio De Simone, “Ushahidi and the Power of Crowdsourcing,” InfoQ, June 27, 2015; Ushahidi homepage, 57Sergio De Simone，"Ushahidi和众包的力量"，InfoQ，2015年6月27日；Ushahidi主页，https://www.
https://www.ushahidi.com/. ushahidi.com/。
28 28
on a “cyber-recon team” report containing two critical pieces of enemy information obtained 在一份"网络侦察小组"报告中，包含完全从facebook，Snapchat和Tinder上的开源信息中
entirely from open-source information on Facebook, Snapchat, and Tinder.58 获得的两个关键信息。 58
DARPA’s Air Space Total Awareness for Rapid Tactical Execution program seeks to develop DARPA的Air Space Total Awareness For Rapid Tactical Execution program旨在开发新的先
new advanced low‐cost sensors, AI algorithms, and virtual testing environments to deconflict 进低成本传感器，AI算法和虚拟测试环境，以消除友军的空域活动并创建更好的共同操作图
airspace activities of friendly forces and create a better common operating picture.59 景。 59
Computer Vision 计算机视觉

Computer Vision (CV), apart from representing one of the earliest60 and still most intensely 计算机视觉（CV）除了代表人工智能和机器学习中最早的60个也是研究最深入的领域之一
researched areas in AI and ML,61serves as a methodological backbone to several powerful 外，61还作为几种强大的衍生技术的方法支柱，这些技术可以用来减轻平民的伤害，包括
基本的物体识别和分类；62运动、行为和步态分析；63和动态环境中的一般感知、学习和
derivative technologies that can be leveraged to mitigate civilian harm, including basic object
跟踪（与机器人技术协同，见上文）。 64
recognition and classification;62 motion, behavior, and gait analysis;63 and general perceiving,
learning, and tracking in dynamic environments (in concert with Robotics, see above).64
58Christopher Lowman and Gerald Prater, “Expansion of the Reconnaissance and Security BCT into the Cyber 58Christopher Lowman和Gerald Prater，"将侦察和安全BCT扩展到网络领域：从NTC轮换中吸取的经验教训1
Domain: Lessons Learned from NTC Rotation 17-07.05,” July 2017, unpublished white paper. During the National 7-07.05，"2017年7月，未发表的白皮书。在全国
Training Center exercise, the best estimate of opposing-force (OPFOR) locations based “entirely on social-media 训练中心演习，对"完全基于社交媒体拖钓......的对手部队（OPFOR）位置的最佳估计与来自其他来源的模板化O
trolling … was surprisingly consistent with templated OPFOR locations derived from other sources,” see Curt PFOR位置惊人地一致，"参见Curt Taylor，"骑兵是时候认真对待网络侦察了，"Armor：Mounted机动杂志（20
Taylor, “It’s Time for Cavalry to Get Serious about Cyber Reconnaissance,” Armor: Mounted Maneuver Journal (Fall 18年秋季），https://mcoe.azurewebsites.us/Armor/eARMOR/content/issues/2018/Fall/4Taylor18.pdf...
2018), https://mcoe.azurewebsites.us/Armor/eARMOR/content/issues/2018/Fall/4Taylor18.pdf.
59DARPA states, “ASTARTE will not only provide a continuously updating, real-time, four-dimensional (space and 59DARPA指出，"ASTARTE不仅将为友军提供不断更新的实时四维（空间和时间）战斗空间运动图像，还将使
time) moving picture of the battlespace for friendly forces but will also use its sensor network to detect and map 用其传感器网络检测和绘制对手位置，提高态势感知能力。"请参阅"未来战斗的实时空域意识和消除冲突，"DA
adversary locations, increasing situational awareness.” See “Real-Time Airspace Awareness and De-Confliction for RPA网站，Apr。 2020年7月https://www.darpa.mil/news-events/2020-04-07...
Future Battles,” DARPA website, Apr. 7, 2020, https://www.darpa.mil/news-events/2020-04-07.
60Although CV has been studied for decades (dating back to the early 1960s), a significant milestone occurred in 60尽管CV已经研究了几十年（可追溯到20世纪60年代初），但2015年2月发生了一个重要的里程碑，当时微
February 2015, when Microsoft Research announced a system that for the first time surpassed human-level 软研究院宣布了一个系统，该系统首次在ImageNet上超越参见Kaiming He et al.，"深入研究整流器：在Image
performance on ImageNet, a widely used dataset. See Kaiming He et al., “Delving Deep into Rectifiers: Surpassing Net分类上超越人类水平的性能，"2月2日。 6，2015，arXiv：1502.01852v1[cs.CV]。
Human-Level Performance on ImageNet Classification,” Feb. 6, 2015, arXiv:1502.01852v1 [cs.CV].
61Joel Janai et al., “Computer Vision for Autonomous Vehicles: Problems, Datasets and State of the Art,” Mar. 17, 61Joel Janai等人。，"自动驾驶汽车的计算机视觉：问题，数据集和艺术状态，"Mar。 17，2021，arXiv：
2021, arXiv:1704.05519v3 [cs.CV]. 1704.05519v3[cs.CV]。
62L.E. Carvalho and A. von Wangenheim, “3D Object Recognition and Classification: A Systematic Literature 62L.E.Carvalho和a.von Wangenheim，"3D对象识别和分类：系统文献综述，"模式分析和应用22（Feb.
Review,” Pattern Analysis and Applications 22(Feb. 2019). 2019).
63Salisu Ibrahim Yusuf, Steve Adeshina, and Moussa Mahamat Boukar, “Parameters for Human Gait Analysis: A 63Salisu Ibrahim Yusuf，Steve Adeshina和Moussa Mahamat Boukar，"人类步态分析的参数：综述"，2019年
Review,” in 2019 15th International Conference on Electronics, Computer and Computation (ICECCO), Dec. 2019. 第15届电子，计算机和计算国际会议（ICECCO），12月。 2019.
64Muhammad Saputra, Andrew Markham, and Niki Trigoni, “Visual SLAM (Simultaneous Localization and 64Muhammad Saputra、Andrew Markham和Niki Trigoni，"动态环境中运动的视觉SLAM（同时定位和映射）
Mapping) and Structure from Motion in Dynamic Environments: A Survey,” ACM Computing Surveys51, no. 2 (June 和结构：调查"，ACM计算调查51，第2期（2018年6月）。
2018).
29 29
A well-known recent example of this kind of technology is Project Maven,65 which consists of 这种技术的一个众所周知的最近的例子是Project Maven，65它包括用于从无人机图像中识
tools for identifying potential targets (e.g., vehicles, buildings, and people) from UAV imagery. 别潜在目标（例如，车辆，建筑物和人员）的工具。此外，DARPA的目标识别和适应在竞
Also, DARPA’s Target Recognition and Adaption in Contested Environments program uses ML 争环境计划使用ML来定位和识别目标。 66
to locate and identify targets.66
Other examples of AI/ML-enabled CV systems or research programs applicable to civilian harm 适用于减轻平民伤害的AI/ML启用CV系统或研究计划的其他示例包括以下内容:
mitigation include the following:
• The Army’s Next-Generation Squad Weapon, which is slated to replace the M249 • 陆军的下一代小队武器，计划取代M249小队自动武器和M4/M4A1卡宾枪，据称将
squad automatic weapon and the M4/M4A1 carbine and will purportedly be equipped 配备自动目标识别，目标跟踪，能够将火控数据传输给其他人的无线通
with automatic target recognition, target tracking, wireless communication able to
transmit fire control data to others, and facial recognition technology67
• Synthetic aperture radar (SAR) target recognition with DL68 • 利用DL68进行合成孔径雷达(SAR)目标识别
• Drone-based thermal imaging and recognition69 and • 基于无人机的热成像和识别69和
• Using Wi-Fi70 and shadows71 to “see” behind walls and around corners, respectively, • 使用Wi-Fi70和阴影71分别"看到"墙壁后面和角落周围，并使用单个激光射击72（
and using a single laser shot72 (fired through a keyhole) to track moving objects 通过钥匙孔发射）跟踪移动物体
65 Project Maven (formally known as the Algorithmic Warfare Cross-Functional Team) was established in 2017 to 65Project Maven（正式名称为算法战跨职能团队）成立于2017年，旨在"加速DOD对大数据和机器学习的整合
“accelerate the DOD’s integration of big data and machine learning.” The ownership of Project Maven (as of this 。"项目Maven的所有权（截至撰写本文时）正在从华盛顿总部服务处转移到负责情报和安全的国防部副部长，
writing) is being transferred from the Washington Headquarters Service to the Under Secretary of Defense for 正如当时的国防部副部长罗伯特*沃克（Robert Work）在原始备忘录中所规定的那样，他的办公室建立了该计
Intelligence & Security, as stipulated in the original memo by then-Deputy Defense Secretary Robert Work, whose 划。参见Robert Work,Memorandum,Subject:Establishment of An Algorithmic Warfare Cross-Functional Team
office established the program. See Robert Work, Memorandum, Subject: Establishment of an Algorithmic Warfare (Project Maven),Apr. 2017年https://dodcio.defense.gov/Portals/0/Documents/Project%20Maven%20DSD%20
Memo%2020170425.pdf...
Cross-Functional Team (Project Maven), Apr. 26, 2017,
https://dodcio.defense.gov/Portals/0/Documents/Project%20Maven%20DSD%20Memo%2020170425.pdf.
66Pat Host, “Deep Learning Analytics Develops DARPA Deep Machine Learning Prototype,” Defense Daily, May 11, 66Pat主持人，"深度学习分析开发DARPA深度机器学习原型"，国防日报，2016年5月11日，https://www.def
2016, https://www.defensedaily.com/deep-learning-analytics-develops-darpa-deep-machine-learning- ensedaily.com/deep-learning-analytics-develops-darpa-deep-machine-learningprototype/advanced-transf
prototype/advanced-transformational-technology/. ormational-technology/.
67The award date is scheduled for November 2021, with delivery of the first rifle/automatic rifle systems to the 67奖励日期定于11月2021，5月2022向军队交付第一支步枪/自动步枪系统。参见Mathew Cox，"陆军的下
Army in May 2022. See Mathew Cox, “Army’s Next Infantry Weapon Could Have Facial-Recognition Technology,” 一个步兵武器可能有面部识别技术，"
Real Clear Defense, June 3, 2019. 真正清晰的防御，2019年6月3日。
Ryan J. Soldin, “SAR Target Recognition with Deep Learning,” in IEEE Applied Imagery Pattern Recognition
68 68Ryan J.Soldin,"Sar Target Recognition with Deep Learning,"in IEEE Applied Imaging Pattern Recognitio
Workshop (AIPR), Washington, DC, Oct. 2018. n Workshop(AIPR),Washington,DC,Oct. 2018.
69David C. Schedl, Indrajit Kurmi, and Oliver Bimber, “Search and Rescue with Airborne Optical Sectioning,” 69David C.Schedl，Indrajit Kurmi和Oliver Bimber，"使用机载光学切片进行搜索和救援"，Nature Machi
Nature Machine Intelligence 2 (Nov. 23, 2020). ne Intelligence2（11月11日）。 23, 2020).
70Mingmin Zhao et al., “Through-Wall Human Pose Estimation Using Radio Signals,” Computer Vision and Pattern 70Mingmin Zhao等。，"使用无线电信号进行穿墙人体姿势估计"，计算机视觉和模式识别（CVPR）（2018
Recognition (CVPR) (2018), http://openaccess.thecvf.com/content_cvpr_2018/papers/Zhao_Through- ），http://openaccess.thecvf.com/content_cvpr_2018/papers/Zhao_Through
Wall_Human_Pose_CVPR_2018_paper.pdf. Wall_Human_Pose_CVPR_2018_paper.pdf.
71Charles Saunders, John Murray-Bruce, and Vivek Goyal, “Computational Periscopy with an Ordinary Digital 71Charles Saunders，John Murray-Bruce和Vivek Goyal，"普通数码相机的计算潜望镜，"Nature365（1
Camera,” Nature 365 (Jan. 23, 2019), https://www.nature.com/articles/s41586-018-0868-6. 月1日）。 23，2019），https://www.nature.com/articles/s41586-018-0868-6...
72Christopher A. Metzler, David B. Lindell, and Gordon Wetzstein, “Keyhole Imaging: Non-Line-of-Sight Imaging 72Christopher A.Metzler,David B.Lindell,And Gordon Wetzstein,"Keyhole Imaging:Non-Line-Of-Sight Imagin
and Tracking of Moving Objects Along a Single Optical Path,” IEEE Transactions on Computational Imaging 7 (Dec. g and Tracking of Moving Objects Along A Single Optical Path,"IEEE Transactions on Computational Imaging7(
22, 2020). Dec. 22, 2020).
30 30
Enabling Technologies 启用技术
Enabling Technologies are a set of miscellaneous AI/ML-enabled technologies (all of which 启用技术是一组杂项AI/ML启用的技术（所有这些都取决于其他应用程序），包括生物识别
depend on other applications) that includes biometric and forensic analysis tools,73 edge 和取证分析工具，73边缘计算（和"智能"传感器），74和多传感器和多域 75赋能技术从（
computing (and “smart” sensors),74 and multi-sensor and multi-domain sensor fusion.75 1）概念上--如"AI堆栈"框架，该框架由一组相互依赖的技术层组成，以帮助可视化、组织
、计划和优先考虑战略AI/ML能力和投资76--到（2）跨学科--如陆军研究办公室资助的一
Enabling technologies run the gamut from (1) conceptual—as in the “AI stack” framework, 个项目，该项目应用认知神经科学的洞察力开发新的"大脑训练"方法，帮助士兵避免友军火
which consists of a set of interdependent technology layers to help visualize, organize, plan, 力对平民的伤害77--到（3）有远见--如在战术边缘的协作AI（CATE），最近开发原型AI/
and prioritize strategic AI/ML capabilities and investments76—to (2) cross-disciplinary—as in ML的努力架构和框架，使协作、多智能人工智能技术能够简单、快速地集成到边缘、传感
an Army Research Office–funded project that applied insights from cognitive neuroscience to 器和战术层面的处理、开发和传播链中。 CATE"想象一名士兵在巡逻......[在]背景中，协作AI
develop new “brain training” methods to help soldiers avoid civilian harm by friendly fire77— 代理扫描城市相机，查看生活模式，提供通过监视启用的AI。人工智能特工确定存在威胁
并提醒士兵，他们永远不会低头看屏幕并将目光从周围环境移开。” 78
to (3) visionary—as in collaborative AI at the tactical edge (CATE), a recent effort to develop a
prototype AI/ML architecture and framework that enables simple, rapid integration of
collaborative, multiagent AI technology into the processing, exploitation, and dissemination
chain at the edge, sensor, and tactical level. CATE “imagines a soldier on patrol … [in] the
background, collaborative AI agents scan city cameras, review patterns of life, providing an AI
enabled over watch. The AI agents determine there is a threat and alert the soldier, who never
[has] to look down at a screen and take his eyes off his immediate surroundings.”78
Knowledge Representation and Reasoning 知识表示与推理

Knowledge Representation and Reasoning (KR&R) constitutes a critical set of nascent 知识表示和推理（kr＆R）构成了一组新兴技术的关键，这些技术正在开发中，以表示有关
technologies that are being developed to represent contextually rich information about the 世界的上下文丰富信息。 79虽然许多更具体的技术和应用可以归入本
world.79 Although many more specific techniques and applications can be subsumed into this
73Mayank Vatsa, Richal Singh, and Angshul Majumdar, eds., Deep Learning in Biometrics, (Boca Raton, FL: CRC 73Mayank Vatsa，Richal Singh和Angshul Majumdar，eds。，Deep Learning in Biometrics，（Boca Rato
Press, 2018). n，FL：CRC Press，2018）。
74 Mario Molinara, Alessandro Bria, Saverio De Vito, and Claudio Marrocco, eds., “Artificial Intelligence for 74Mario Molinara，Alessandro Bria，Saverio De Vito和Claudio Marrocco，eds。，"人工智能的分布
Distributed Smart Sensing,” special issue, Pattern Recognition Letters (Jan. 2021). 式智能传感，"特刊，模式识别信（1月1日）。 2021).
75Erik Blasch et al., “Machine Learning/Artificial Intelligence for Sensor Data Fusion–Opportunities and 75Erik Blasch等人。，"用于传感器数据融合的机器学习/人工智能-机遇与挑战"，IEEE航空航天和电
Challenges,” IEEE Aerospace and Electronic Systems Magazine 36, no. 7 (July 2021). 子系统杂志36，第7期（2021年7月）。
76Andrew Moore, Martial Hebert, and Shane Shaneman, “The AI Stack: A Blueprint for Developing and Deploying 76Andrew Moore、Martial Hebert和Shane Shaneman，"AI堆栈：开发和部署人工智能的蓝图"，探讨持久性IS
Artificial Intelligence,” in Ground/Air Multisensor Interoperability, Integration, and Networking for Persistent ISR R的地面/空中多传感器互操作性、集成和网络
IX, Proceedings Vol. 10635, May 4, 2018. The lead author of this paper served as dean of the Carnegie Mellon 第九卷。 10635，2018年5月4日。本文的主要作者曾担任卡内基梅隆大学院长
University School of Computer Science from 2014 to 2018 and was a member of the NSCAI. 大学计算机科学学院从2014年到2018年，是NSCAI的成员。
77Adam Biggs, Matthew Cain, and Stephen Mitroff, “Cognitive Training Can Reduce Civilian Casualties in a 77Adam Biggs，Matthew Cain和Stephen Mitroff，"认知训练可以在模拟射击环境中减少平民伤亡，"
Simulated Shooting Environment,” Psychological Science (July 13, 2015). 心理科学（2015年7月13日）。
78Susan Toth and William Hughes, “The Journey to Collaborative AI at the Tactical Edge (CATE),” in Artificial 78Susan Toth和William Hughes,"The Journey to Collaborative AI at The Tactical Edge(CATE),"In Artificial Intelli
Intelligence and Machine Learning for Multi-Domain Operations Applications III, Proceedings Vol. 11746, Apr. 12, gence and Machine Learning for Multi-Domain Operations Applications III,Proceedings Vol. 11746年4月 12, 2
2021. 021.
79 Pierre Marquis, Odile Papini, and Henri Prade, eds., A Guided Tour of Artificial Intelligence Research: Volume I: 79Pierre Marquis，Odile Papini和Henri Prade，eds。，人工智能研究的导游：第一卷：知识表示，推理和
Knowledge Representation, Reasoning and Learning, (Springer-Verlag, May 2020). 学习，（Springer-Verlag，2020年5月）。
31 31
one category, Figure 3 identifies three areas that can most readily be leveraged to help mitigate 一个类别，图3确定了三个最容易被用来帮助减轻平民伤害的领域：语义网络和本体，80因
civilian harm: semantic networks and ontologies,80 causal reasoning and discovery,81 and 果推理和发现，81和推理推理。 82
inferential reasoning.82
An example of a large-scale KR&R system (albeit one that is still at the basic research front) is 大规模KR＆R系统的一个例子（尽管仍然处于基础研究前沿）是DARPA通过规划活动情境
DARPA’s Collection and Monitoring via Planning for Active Situational Scenarios (COMPASS) 情景（COMPASS）计划的收集和监控。指南针的设计是利用人工智能、博弈论、建模和估
program. COMPASS is designed to leverage AI, game theory, and modeling and estimation to 计来识别刺激，这些刺激能产生关于对手意图的最多信息，并为决策者提供关于如何做出反
应的高保真情报（对每个行动过程进行积极和消极的权衡）。 83
both identify stimuli that yield the most information about an adversary’s intentions and
provide decision-makers high-fidelity intelligence on how to respond (with positive and
negative tradeoffs for each course of action).83
AI/ML-enabled knowledge representation tools can also be used to mitigate common shortfalls 支持AI/ML的知识表示工具也可用于减轻"任务切换"的常见不足，"任务切换"定义为"将正
of “mission handoff,” defined as the “process of passing an ongoing mission from one unit to 在进行的任务从一个单位传递到另一个单位而没有明显丧失连续性的过程"（例如SA，对手
another with no discernible loss of continuity” (e.g., SA, adversary composition, allies, host 组成，盟友，东道国部队，平民）。 84典型的问题包括切换时间周期短、操作环境的动态
性、传出单位提供的数据与传入单位所需的数据之间的不匹配以及格式上的不匹配(例如在
nation forces, civilian populace).84 Typical problems include short handoff timeframes, the 情报简报中)。 AI/ML基于知识的查询工具（结合"理解"非结构化文本数据的方法，请参阅
dynamic nature of the operational environment, mismatches between data offered by an 下面的自然语言处理应用程序）可用于深入了解所有相关数据。 85
outgoing unit and data required by an incoming unit, and formatting mismatches (e.g., in
intelligence briefs). AI/ML knowledge-based query tools (combined with methods to
“understand” unstructured textual data, see Natural Language Processing applications, below)
can be used to gain insight into all relevant data.85
Natural Language Processing 自然语言处理

Natural Language Processing (NLP), like CV, is a catchall phrase that consists of a vast set of 自然语言处理（NLP）和CV一样，是一个集文本自动操作为中心的大量相互关联的方法和
interrelated methodologies and applications centered on the automatic manipulation of text- 应用于一体的流行语
80 Ji Han et al., “Semantic Networks for Engineering Design: State of the Art and Future Directions,” Journal of 80季汉等。，"工程设计的语义网络：艺术现状和未来方向"，机械设计杂志144，第2期（9月）。 9, 202
Mechanical Design 144, no. 2 (Sept. 9, 2021). 1).
81Judea Pearl, “The Seven Tools of Causal Inference, with Reflections on Machine Learning,” Communications of the 朱迪亚*珀尔，"因果推理的七个工具，对机器学习的反思"，ACM62通讯，第3期（3月）。 2019).
ACM 62, no. 3 (Mar. 2019).
82Sagir Yusuf and Chris Baber, “Inferential Reasoning for Heterogeneous Multi-Agent Mission,” International 82Sagir Yusuf和Chris Baber，"异构多智能体任务的推理"，国际学术和科学研究与创新14，第10期（202
Scholarly and Scientific Research & Innovation 14, no. 10 (2020). 0年）。
83Fotis Barlos et al., Collection and Monitoring via Planning for Active Situational Scenarios (COMPASS): Strategic 83Fotis Barlos等人。，收集和监测通过规划活动情境情景（指南针）：战略多层评估（SMA）报告，桑迪亚
Multi-Layer Assessment (SMA) Report, Sandia National Laboratories, SAND2020-0136R, Jan. 2020, 国家实验室，SAND2020-0136r，1月。 2020年，https://www.osti.gov/servlets/purl/1592839...
https://www.osti.gov/servlets/purl/1592839.
84Department of Defense, Foreign Internal Defense, Joint Publication 3-22, Aug. 17, 2018, 84国防部，外国内部防御，联合出版物3-22，8月。 2018年17月https://www.jcs.mil/
https://www.jcs.mil/Portals/36/Documents/Doctrine/pubs/jp3_22.pdf. Portals/36/Documents/Doctrine/pubs/jp3_22.pdf...
85 See pages 89–90 in Stephan De Spiegeleire, Matthijs Maas, and Tim Sweijs, Artificial Intelligence and the Future 85见Stephan De Spiegeleire,Matthijs Maas,And Tim Sweijs,Artificial Intelligence and The Future Of Defense:S
of Defense: Strategic Implications for Small and Medium-Sized Force Providers, The Hague Centre for Strategic trategic Implications for Small And Medium-Sized Force Providers,The Hague Centre For Strategic Studies,20
Studies, 2017, https://www.jstor.org/stable/resrep12564. 17,https://www.jstor.org/stable/resrep12564...
32 32
and speech-based language.86 Specific NLP tasks (for which there are numerous extant ML 和基于语音的语言。 86具体的NLP任务（其中有许多现存的ML应用）包括机器翻译，87文
applications) include machine translation,87 text classification and summarization,88 semantic 本分类和摘要，88语义解析，89情感分析，90语音到文本转录，91和自然语言推理和知识
parsing,89 sentiment analysis,90 speech-to-text transcription,91 and natural language inference 提取。 92此外，NLP方法可用于开发培养文化意识的培训工具。 93至于几乎所有其他情况
图3中的条目仅代表一组大得多的NLP应用程序的一小部分说明性子集这些应用程序可用于
and knowledge extraction.92 Additionally, NLP methods can be used to develop training tools 帮助减轻平民伤害。
to foster cultural awareness.93 As for almost all other cases, the entries in Figure 3 represent
only a small illustrative subset of a vastly larger set of NLP applications that may be leveraged
to help mitigate civilian harm.
Planning and Scheduling 计划及时间表

Planning and Scheduling (P&S) represents a core set of “classic AI” methods that consist of 计划和调度（p&S）代表了一套核心的"经典人工智能"方法，包括决定在复杂的动态和不确
deciding on a course of action and steps to take in complex dynamic and uncertain 定的情况下采取的行动和步骤
86Amirsina Torfi et al., “Natural Language Processing Advancements by Deep Learning: A Survey,” Feb. 27, 202, 86Amirsina Torfi等人。，"自然语言处理深入学习的进步：一项调查，"2月2日。 27，202，arXiv：2003.01
arXiv:2003.01200v4 [cs.CL]. 200v4[cs.CL]。
87Darminder Ghataoura and Sam Ogbannaya, “Application of Image Captioning and Retrieval to Support Military 87Darminder Ghataoura和Sam Ogbannaya，"图像字幕和检索支持军事决策的应用"，in International Confere
Decision Making,” in International Conference on Military Communication and Information Systems (ICMCIS), the nce on Military Communication And Information Systems(ICMCIS),the
Hague, Netherlands, May 4-5, 2021, IEEE. 荷兰海牙，2021年5月4-5日，IEEE。
88Shenguluan Hou and Ruqian Lu, “Knowledge-Guided Unsupervised Rhetorical Parsing for Text Summarization,” 88Shenguluan Hou And Ruqian Lu,"Knowledge-Guided Unsupervised修辞解析for Text Summarization,"Infor
Information Systems 94 (Dec. 2020). mation Systems94(Dec. 2020).
Hossam Elzayady, Khaled Badran, and Gouda Salama, “Arabic Opinion Mining Using Combined CNN - LSTM
89 89Hossam Elzayady，Khaled Badran和Gouda Salama，"使用联合CNN LSTM模型进行阿拉伯意见挖掘"，
Models,” International Journal of Intelligent Systems and Applications 12, no. 4 (Aug. 2020.) 《国际智能系统与应用杂志》12，第4期（8月8日）。 2020.)
90 Recent work on applying ML and DL to sentiment analysis (with relevance to military applications) includes (1) 90最近关于将ML和DL应用于情感分析（与军事应用相关）的工作包括（1）Liang-Chu Chen，Chia-Meng Lee
Liang-Chu Chen, Chia-Meng Lee, and Mu-Yen Chen, “Exploration of Social Media for Sentiment Analysis Using Deep 和Mu-Yen Chen，"Exploration of Social Media For Sentiment Analysis Using Deep
Learning,” Soft Computing 24 (2020); (2) Mohammed El-Jawad, Rania Hodlod, and Yasser Omar, “Sentiment 学习，"软计算24（2020）;（2）Mohammed El-Jawad，Rania Hodlod和Yasser Omar，"。
Analysis of Social Media Networks Using Machine Learning,” in 14th International Computer Engineering 使用机器学习的社交媒体网络分析，"在第14届国际计算机工程
Conference, Cairo, Egypt, Dec. 29-30, 2018, IEEE; and (3) Yogesh Chandra and Antoreep Jana, “Sentiment Analysis 会议,开罗,埃及,12月12日。 2018年29-30日，IEEE；以及（3）Yogesh Chandra和Antoreep Jana，"使用机器
Using Machine Learning and Deep Learning,” in 7th International Conference on Computing for Sustainable Global 学习和深入学习的情感分析"，在第七届可持续全球计算国际会议上
Development, New Delhi, India, Mar. 12-14, 2020, IEEE. 发展，新德里，印度，3月。 12-14,2020,IEEE.
91 Abdelaziz Abdelhamid et al., “End-to-End Arabic Speech Recognition: A Review,” in 19th Conference on 91Abdelaziz Abdelhamid等人。，"端到端阿拉伯语语音识别：回顾"，在第19届语言工程会议（ESOLE
Language Engineering (ESOLEC’19), Sept. 2020. C'19），9月。 2020.
92Robert E. Wray, James Kirk, and John Laird, “Language Models as a Knowledge Source for Cognitive Agents,” 92Robert E.Wray，James Kirk和John Laird，"语言模型作为认知代理的知识来源"，9月9日。 20，2021，a
Sept. 20, 2021, arXiv:2109.08270v2 [cs.AI]. rXiv：2109.08270v2[cs.AI]。
93Sodiq Adewole et al., “Dialogue-Based Simulation for Cultural Awareness Training,” Feb. 1, 2020, 93Sodiq Adewole等人。，"基于对话的文化意识培训模拟"，2月2日。 1，2020，arXiv：2002.
arXiv:2002.00223v1 [cs.CY]. 00223v1[cs.CY]。
33 33
environments.94 Basic civilian harm-centric methods include planning under uncertainty,95 环境。 94以平民危害为中心的基本方法包括不确定性下的规划95多代理规划和任务96和动
multi-agent planning and tasking,96 and real-time response to dynamic environments.97 态环境的实时响应。 97
The use of AI for dynamic planning (at least pertaining to optimizing and scheduling military 人工智能用于动态规划（至少与优化和调度军事物流问题有关）的使用可以追溯到DARPA
logistics problems) goes back several decades to DARPA’s Dynamic Analysis and Replanning 的动态分析和再规划工具程序几十年前，该程序于1991年推出，用于在沙漠风暴行动期间
Tool program, introduced in 1991 and used for planning logistics during Operation Desert 规划物流。 98其他例子包括AFRANCI（2006年开发），它将神经网络建模与符号算法相结
合，并已被用作支持民用救援情景的P＆S工具，99和协调的连续蒙特卡洛树搜索算法，已
Storm.98 Other examples include AFRANCI (developed in 2006), which combines neural 100（见几个与战争游戏相关的行动过程示例的培训。)
network modeling with symbolic algorithms and has been used as a P&S tool to support a
civilian rescue scenario,99 and a coordinated continuous Monte Carlo tree-search algorithm,
which has been applied to planning search and rescue missions for UAV teams.100 (See Training
for several wargaming-related course of action examples.)
Decision-Making Decision-Making
Decision-Making overlaps with both KR&R and P&S (to the extent that all three application 决策与KR&R和P&S重叠（因为所有三个应用领域都涉及旨在裁决行动方针的方法），并包
areas entail methods designed to adjudicate a course of action) and includes applications at the 括在更广泛的知识表示/规划/决策过程的"决策"端应用。 101以伤害为中心的方法包括那些
“decision” end of the broader knowledge representation/planning/decision-making 可用于探索新颖
process.101 civilian harm-centric methods include those that can be used to explore novel
94See Chapter 11 in Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th Edition, 94见Stuart Russell和Peter Norvig《人工智能：现代方法》第4版Pearson2021年5月第11章。
(Pearson, May 2021).
95Yang Zhen, Zhang Wanpeng, and Liu Hongfu, “Real-Time Strategy Game Tactical Recommendation Based on 95杨震，张万鹏，刘洪福，《基于贝叶斯网络的即时战略游戏战术推荐》，《物理学报：会议辑1168》，
Bayesian Network,” Journal of Physics: Conference Series 1168, no. 3 (2019). 第3期（2019年）。
96Alejandro Torreno et al., “Cooperative Multi-Agent Planning: A Survey,” ACM Computing Surveys 50, no. 6 (Nov. 96Alejandro Torreno等人。，"合作多代理规划：一项调查"，ACM计算调查50，第6期（11月11日）。 2017
2017). ).
97Chao Chen et al., “NECTAR-An Agent-Based Dynamic Task Allocation Algorithm in the UAV Swarm,” Complexity 97Chao Chen等人。，"花蜜-无人机群中基于代理的动态任务分配算法，"复杂性2020（9月）。 2020).
2020 (Sept. 2020).
98 Sara Reese Hedberg, “DART: Revolutionizing Logistics Planning,” IEEE Intelligent Systems 17 (May 2002). 98Sara Reese Hedberg,"DART:Revolutionizing Logistics Planning,"IEEE Intelligent Systems17(May2002).
99 Specifically, the testbed problem was to use the system to “decide” whether an ambulance or firefighter should 99具体而言，testbed问题是使用系统来"决定"救护车或消防员是否应该拯救平民。假设平民在燃烧的建筑物
rescue a civilian. The civilian is assumed to be somewhere in a burning building, and the decision is based on an 的某个地方，决定是基于"代理人"的位置和平民的生活状况。动态变量可以包括救护车、消防员、平民、燃烧
“agent’s” location and the civilian’s life condition. Dynamic variables may include the positions of the ambulance, 建筑物、消防队和最近的避难所的位置；描述消防员和平民的"生活状况"的度量；建筑物的波动程度；以及执
firefighter, civilian, burning building, fire brigade, and nearest refuge; metrics describing the “life condition” of the 行平民类似CIVCAS相关的"问题"的类比应该是显而易见的。参见Francisco Reinaldo等人。，"多策略学习的工
firefighter and civilian; the building’s degree of volatility; and the innate difficulty of performing the civilian 具"计算科学AI研究进展26（2006）。
rescue. The analogy to similar CIVCAS-related “problems” ought to be obvious. See Francisco Reinaldo et al., “A
Tool for Multi-Strategy Learning,” Advances in AI Research in Computing Science 26 (2006).
100Chris Baker et al., “Planning Search and Rescue Missions for UAV Teams,” in Proceedings of the Twenty-second 100克里斯贝克等人。，"规划无人机队的搜索和救援任务，"在第二十二届欧洲人工智能会议记录，8月8日。
European Conference on Artificial Intelligence, Aug. 2016. 2016.
Yash Shrestha, Shiko Ben-Mehahem, and Georg von Krogh, “Organizational Decision-Making Structures in the
101 101Yash Shrestha，Shiko Ben-Mehahem和Georg von Krogh，"人工智能时代的组织决策结构"，《加州管理
Age of Artificial Intelligence,” California Management Review 61, no. 4 (Aug. 2019). 评论》61，第4期（8月15日）。 2019).
34 34
tactics and strategies in abstract military operational mission and policy spaces102 and perform 战术和战略在抽象的军事行动任务和政策空间102和执行多维风险评估103和那些严格建立
multi-dimensional risk assessments103 and those that are rigorously grounded in a system-of- 在系统系统方法的一般设计过程。 104
systems approach to design processes in general.104
The most notable recent successes of AI/ML-enabled decision-making (outside of DOD) have AI/ML支持的决策（DOD之外）最近最显着的成功来自对游戏玩算法的研究。示例包括以
come from research into game-playing algorithms. Examples include the following: 下内容:
• AlphaZero, which, starting from random play and using no domain knowledge except • AlphaZero从随机游戏开始，除了游戏规则外，不使用任何领域知识，只需要24小
for game rules, required only 24 hours to achieve a superhuman level of play in chess, 时就能在国际象棋、将棋（日本的国际象棋变体）和围棋中达到超人的水平，并在
shogi (a Japanese variant of chess), and Go, and defeated a world champion program 每个105年击败了一个世界冠军计划（后来的版本MuZero在没有任何游戏规则知识
的情况下与AlphaZero的超人表现相匹配）
in each105 (a later version, MuZero, matched AlphaZero's superhuman performance
without any knowledge of game rules)106
• An AI that learned to play all 57 Atari video games107 • 一个学会玩所有57Atari视频游戏的AI107
• AlphaStar, which defeated 99.8 percent of human Starcraft II gamers108 • AlphaStar，它击败了99.8％的人类星际争霸ii游戏玩家108
• Pluribus, the first AI to defeat human professional players in a multiplayer game109 • Pluribus，第一个在多人游戏中击败人类职业玩家的AI109
An important implicit additional dimension that can be used to characterize all AI-enabled 可用于表征所有AI支持的决策过程的一个重要的隐含附加维度是聚合的人-AI决策关系。 1
decision-making processes is the aggregated human-AI decision-making relationship.110 For 10为
102Bonnie Johnson and William Treadway, “Artificial Intelligence—An Enabler of Naval Tactical Decision 102Bonnie Johnson和William Treadway，"人工智能-海军战术决策优势的推动者"，AI杂志40，no.1
Superiority,” AI Magazine 40, no. 1 (Mar. 2019). （Mar。 2019).
103Nicola Paltrinieri, Louise Comfort, and Genserik Reniers, “Learning About Risk: Machine Learning for Risk 103Nicola Paltrinieri，Louise Comfort和Genserik Reniers，"学习风险：用于风险评估的机器学习"，安全
Assessment,” Safety Science 118 (Oct. 2019). 科学118（10月11日）。 2019).
104Kanstantsin Miatliuk, Adam Wolniakowski, and Pawel Kolosowski, “Engineering System of Systems Conceptual 104Kanstantsin Miatliuk，Adam Wolniakowski和Pawel Kolosowski，"分层系统理论基础中的系统概念设计工
Design in Theoretical Basis of Hierarchical Systems,” in 2020 IEEE International Conference on Systems, Man, and 程系统"，2020年IEEE系统国际会议，Man和
Cybernetics (SMC), Toronto, Ontario, Canada, Oct. 11–14, 2020, IEEE. 控制论(SMC),多伦多,安大略省,加拿大,10月. 11-14,2020,IEEE.
David Silver et al., “Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm,”
105 105David Silver等人。，"用一般强化学习算法自学掌握国际象棋和将棋，"12月12日。 5，2017，arXiv：171
Dec. 5, 2017, arXiv:1712.01815v1. 2.01815v1。
Julian Schrittwieser et al., “Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model,” Nov. 19,
106 106Julian Schrittwieser等人。，"掌握雅达利，围棋，国际象棋和将棋计划与一个学习的模型，"11月11日
2019, https://arxiv.org/abs/1911.08265. 。 19，2019，https://arxiv.org/abs/1911.08265...
107 Adria Badia et al., “Agent57: Outperforming the Atari Human Benchmark,” Mar. 30, 2020, arXiv:2003.13350. 107Adria Badia等人。，"Agent57：表现优于Atari人类基准，"Mar。 30，2020，arXiv：2003.13350。
108Compared to classic board games like chess, StarCraft II entails much greater real-world-like complexity. For 108与国际象棋等经典棋盘游戏相比，星际争霸II需要更大的真实世界般的复杂性。例如，游戏包括数百个"
example, the game includes hundreds of “pieces” (soldiers in the factions' armies) that move simultaneously in 棋子"（派系军队中的士兵），它们实时同时移动，而不是以有序的回合制方式移动。参见Oriol Vinyals等人
real time, not in an orderly turn‐based fashion. See Oriol Vinyals et al., “Grandmaster Level in StarCraft II Using 。，"星际争霸II中使用多代理强化学习的大师级"，自然575（10月10日）。 30，2019），https://www.nat
Multi-Agent Reinforcement Learning,” Nature 575 (Oct. 30, 2019), https://www.nature.com/articles/s41586- ure.com/articles/s41586019-1724-z...
019-1724-z.
109Unlike chess, in which two players always have perfect access to all game-relevant information, poker includes 109与国际象棋不同，两个玩家总是可以完美地访问所有与游戏相关的信息，扑克包括多个同时进行的玩家，
multiple simultaneous players, and decisions must be based on imperfect (i.e., hidden) information and human- 决策必须基于不完美（即隐藏）的信息和包括虚张声势在内的以人类心理学为中心的行动。参见Noam Brown
psychology-centric actions that include bluffing. See Noam Brown and Tuomas Sandholm, “Superhuman AI for 和Tuomas Sandholm，"Superhuman AI for Multiplayer Poker，"Science365（Aug. 30, 2019).
Multiplayer Poker,” Science 365 (Aug. 30, 2019).
Dominik Dellerman et al., “The Future of Human-AI Collaboration: A Taxonomy of Design Knowledge for
110 110Dominik Dellerman等人。，"人类-人工智能协作的未来：混合智能系统的设计知识分类法"，2021
Hybrid Intelligence Systems,” May 7, 2021, arXiv:2105.03354 [cs.AI]. 年5月7日，arXiv：2105.03354[cs.AI]。
35 35
example, the Sensing for Asset Protection with Integrated Electronic Networked Technology 例如，由英国国防科技实验室开发的具有集成电子联网技术系统的资产保护传感被设计为
system, developed by the UK’s Defence Science and Technology Laboratory, is designed as an 一种人工智能决策支持工具，它将自主传感与融合和传感器管理相结合，提供大面积的SA
AI-enabled decision-support tool that combines autonomous sensing with fusion and sensor 和上下文驱动的决策。 111
management to provide both SA over large areas and context-driven decision-making.111
Targeting Targeting
Targeting denotes a class of AI/ML-enabled applications that combine AI’s innate ability to 定位指的是一类人工智能/ML应用程序，它将人工智能在极高维抽象数据空间中发现模式的
discover patterns in extremely high-dimensional abstract data spaces (see Threat Monitoring, 固有能力与复杂系统理论（CST）112和基于代理的建模（ABM）113社区开发的多个相互
below) with methods of adjudicating multiple conflicting criteria “action selection” problems 冲突的标准"行动选择"问题的方法结合起来。这些方法可以应用于整个F2T2EA靶向周期（
查找，修复，跟踪，目标，参与和评估），114但最以平民伤害为中心的方法是那些包括CS
developed by the complex system theory (CST)112 and agent-based modeling (ABM)113 T/ABM启用的方法来检测，分类，认整体目标选择和损害评估(即图3中主要标题下的另外
communities. These methods can be applied to the entire F2T2EA targeting cycle (find, fix, 两个应用)是指人工智能/ML支持的方法,这些方法尊重相互依存的动态关系,这些关系定义了
track, target, engage, and assess),114 but the most civilian harm-centric ones are those that 一个特定的社会-文化-政治宗教和物理系统作为攻击目标或(攻击后)损害评估的目标。
include CST/ABM-enabled approaches to detect, classify, recognize, and identify potential
targets. Holistic target selection and damage assessment (i.e., the two other applications in
Figure 3 that appear under the main heading Targeting) refer to AI/ML-enabled methods that
respect the interdependent dynamic relationships that define a given socio-cultural-political-
religious and physical system being targeted for attack or assessed for damage (after an
attack).
To illustrate what we mean, at least intuitively, think of the complex food webs in natural 为了说明我们的意思，至少直观地说，想想自然生态中的复杂食物网（大自然自己最好的"
ecologies (nature’s own best exemplars of “complex adaptive systems”). What are the “most 复杂适应性系统"范例）。在给定的生态学中，"最重要"的物种是什么？回答这个问题的一
important” species in a given ecology? One approach to answering this question is to apply 种方法是应用Google的PageRank算法，115对网页的"重要性"进行排名，以确定哪些物种
对于维持生态中的生物生态位至关重要。该方法有效地将网站映射到"物种"（被认为是食
Google’s PageRank algorithm,115 which ranks the “importance” of webpages, to determine 物网的"节点"），并使用通用形式的PageRank来识别其个体损失（食物）的关键物种
which species are critical for sustaining biological niches in the ecology. 116 The method
effectively maps websites to “species” (thought of as “nodes” of a food web) and uses a
generalized form of PageRank to identify the key species whose individual loss (to the food
111UK Ministry of Defence, SAPIENT Middleware Interface Control Document,” May 11, 2020, 111英国国防部，SAPIENT中间件接口控制文件，"2020年5月11日，https://assets.publishing.service.gov.uk/g
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_data/file/895246/S overnment/uploads/system/uploads/attachment_data/file/895246/S APIENT_Interface_Control_Document_
APIENT_Interface_Control_Document_v5.0.pdf. v5.0。pdf。
George Mobus and Michael Kalton, Principles of Systems Science (Understanding Complex Systems), (Springer,
112 112George Mobus和Michael Kalton，系统科学原理（理解复杂系统），（Springer，2016）。
2016).
113 Uri Wilensky and William Rand, An Introduction to Agent-Based Modeling, (Cambridge, MA: MIT Press, 2015). 113Uri Wilensky和William Rand,An Introduction to Agent-Based Modeling,(Cambridge,MA:MIT Press,2015)。
114Merel A. C. Ekelhof, “Lifting the Fog of Targeting: ‘Autonomous Weapons’ and Human Control Through the Lens 114Merel A.C.Ekelhof，"解除目标的迷雾：通过军事目标镜头的'自主武器'和人类控制"，《海军战争学院评论
of Military Targeting,” Naval War College Review 71, no. 3 (2018). 》71，第3期（2018年）。
Amy Langville and Carl Meyer, Google's PageRank and Beyond, (Princeton, NJ: Princeton University Press,
115 115Amy Langville和Carl Meyer，Google'S PageRank And Beyond，（Princeton，NJ：Princeton Univers
2006). ity Press，2006）。
116Stefano Allesina and Mercedes Pascual, “Googling Food Webs: Can an Eigenvector Measure Species' Importance 116Stefano Allesina和Mercedes Pascual，"谷歌搜索食物网：特征向量可以测量物种对共存的重要性吗？"PLoS
for Coextinctions?” PLoS Computational Biology 5, no. 9 (2009). 计算生物学5，no.9（2009）。
36 36
web) would result in the maximal number of other extinctions (i.e., the “most important” web）将导致其他物种灭绝的最大数量（即生态系统中"最重要"的物种）。虽然人们早就
species in an ecosystem). Although it has long been known that the collapse of ecosystems may 知道生态系统的崩溃可能是由关键物种的灭绝引起的，即使是"小"物种的丧失也可能导致多
be triggered by the extinction of critical species117 and that even a “small” species loss may lead 重级联共存，但物种（及其环境）之间通常巨大的相互相互作用网络使得难以预测生态系统
的哪些组成部分是最重要的。从这个简单的例子的关键外卖是，有ai/ML启用的工具（比Pa
to multiple cascading coextinctions,118 the typically vast web of mutual interactions among geRank要复杂得多），使我们能够类比地确定哪些目标，如果受到攻击，最有可能造成平
species (and their environment) makes it difficult to predict which components of an 民伤害。
ecosystem are the most important. The key takeaway from this simple example is that there
are AI/ML-enabled tools (considerably more sophisticated than PageRank) that allow us to
analogously determine which targets, if attacked, are most likely to entail civilian harm.
DARPA’s AI-enabled Mosaic Warfare program offers an explicitly “complex systems of DARPA的人工智能马赛克战争计划提供了一个明确的"复杂系统系统"为基础的作战方法，
systems”-based approach to warfighting designed “around ‘tiles’ of capabilities (i.e., functions: 设计"围绕"能力"（即功能：传感器和射手），而不是独特形状的"拼图"（即平台），必须
安装在战斗计划的特定插槽中才能工作。"119尽管马赛克战争可能距离部署还有几年的时
sensors and shooters), rather than uniquely shaped ‘puzzle pieces’ (i.e., platforms) that must
间，即使是名义形式，但其核心"杀死网络，而不是杀死链"愿景使其特别充满了减轻平民伤
be fitted into specific slots in battle plan in order for it to work.”119 Although Mosaic Warfare 害的机会，如果这些
may still be years away from deployment, even in notional form, its core “kill web, not kill
chain” vision makes it particularly replete with opportunities for civilian harm mitigation if
those opportunities are intentionally built into the design.
Finally, just as for Decision-Making, an important implicit additional class of AI/ML-enabled 最后，就像决策一样，支持AI/ML的定位应用程序的一个重要的隐含附加类是利用混合人-
targeting applications is applications that leverage a hybrid human-AI collaborative effort.120 AI协作努力的应用程序。 120美国陆军的辅助威胁识别从移动合作和自主传感器（ATR MCA
The US Army’s Aided Threat Recognition from Mobile Cooperative and Autonomous Sensors S）系统提供了在短期到中期的"可能的艺术"的一瞥。 ATR-MCAS原型是一个人工智能系统
，由利用传感器和边缘计算的网络化最先进的空中和地面车辆组成。 ATR-MCAS作为士兵
(ATR‐MCAS) system provides a glimpse of the “art of the possible” in the near to medium 的"战场队友"明确开发，包括传感器，使车辆能够在感兴趣的区域内导航，以识别，分类和
term.121 The ATR‐MCAS prototype is an AI‐enabled system that consists of networked state‐of‐ 地理定位实体，障碍物和潜在威胁。它能够聚合和分发目标数据，这些数据可用于根据组
the‐art air and ground vehicles that leverage sensors and edge computing. Being developed 合的威胁图片提出建议和预测，其"支持AI的决策支持代理"建议响应（例如，优先考虑哪些
explicitly as a “battlefield teammate” to soldiers, ATR‐MCAS includes sensors enabling vehicles 威胁）。
to navigate within areas of interest to identify, classify, and geolocate entities, obstacles, and
potential threats. It is capable of aggregating and distributing target data, which can be used to
make recommendations and predictions based on the combined threat picture, and its “AI‐
enabled decision support agent” recommends responses (e.g., which threats to prioritize).
Although only recently announced (in January 2020), DARPA’s Artificial Social Intelligence for 尽管最近才宣布（2020年1月），但DARPA的成功团队人工智能计划旨在开发由具有
Successful Teams program seeks to develop AI/ML-enabled systems built from agents with the
117 Jose Montoya, Stuart Pimm, and Ricard Sole, “Ecological Networks and Their Fragility,” Nature 442 (2006). 117Jose Montoya，Stuart Pimm和Ricard Sole，"生态网络及其脆弱性"，Nature442（2006）。
118 Jose Montoya and Ricard Sole, “Small World Patterns in Food Webs,” Journal of Theoretical Biology 214 (2002). 118Jose Montoya和Ricard Sole，"食物网中的小世界模式"，理论生物学杂志214（2002）。
119“DARPA Tiles Together a Vision of Mosaic Warfare,” DARPA website, Sept. 24, 2018, 119"DARPA Tiles Together A Vision Of Mosaic Warfare，"DARPA网站，9月。 2018年2月24日ht
tps://www.darpa.mil/work-with-us/darpa-tiles-together-a-vision-of-mosiac-warfare...
https://www.darpa.mil/work-with-us/darpa-tiles-together-a-vision-of-mosiac-warfare.
120Jason Cody, Karina Roundtree, and Julie Adams. “Human-Collective Collaborative Target Selection,” ACM 120Jason Cody，Karina Roundtree和Julie Adams。 "人-集体协作目标选择"，ACM Transactions on Hu
Transactions on Human-Robot Interaction 10, no. 2 (Mar. 2021). man-Robot Interaction10，no.2（Mar。 2021).
121Patrick Ferraris, “Aided Detection on the Future Battlefield,” Defense Visual Information Distribution Service, 121Patrick Ferraris，"未来战场上的辅助检测"，国防视觉信息分发服务，1月1日。 2020年https://www.dvid
Jan. 24, 2020, https://www.dvidshub.net/news/360225/aided-detection-future-battlefield. shub.net/news/360225/aided-detection-future-battlefield...
37 37
ability to create shared mental models that “…demonstrate the basic machine social skills 能够创建共享的心理模型，"......展示推断人类合作伙伴的目标和情境知识所需的基本机器社
needed to infer the goals and situational knowledge of human partners, predict what they will 交技能，预测他们将需要什么，并提供上下文感知的行动，以便” 122
need, and offer context‐aware actions in order to perform as adaptable and resilient AI
teammates.”122
Threat Monitoring 威胁监控

Threat Monitoring leverages a broad set of AI/ML-enabled pattern recognition methods and 威胁监控利用了广泛的ai/ML模式识别方法和应用程序。与人类相比，人工智能可以在更短
applications. Compared to humans, AI can assimilate and find patterns in vastly larger data 的时间尺度上吸收和发现更大的数据空间中的模式。 CPLC特定的应用包括行为识别、123
spaces on much shorter time scales. CPLC-specific applications include behavior recognition,123 潜在模式发现、124和将多域数据梦（参见启用技术）转化为可操作的智能。 125
latent pattern discovery,124 and transforming multi-domain datastreams (see Enabling
Technologies) into actionable intelligence.125
Training Training
Training, in the context of mitigating civilian harm, subsumes and conflates myriad other 在减轻平民伤害的背景下，培训将无数其他与AI/ML相关的建模和模拟应用合并并混为一谈
AI/ML-related modeling and simulation applications and includes novel exercise design and ，包括新颖的锻炼设计和
122“ASIST agents must operate in increasingly complex and specialized environments; be adaptable to sudden 122"ASIST特工必须在日益复杂和专业化的环境中工作;适应任务或团队中的突然扰动，如与关键队友失去沟
perturbations in the mission or team, like the loss of communication with a key teammate; and use noisy multi‐ 通;并使用嘈杂的多通道观测来表示世界并做复"参见Joshua Elliot，"成功团队的人工社交智能（ASIST），"D
channel observations to represent the world and do complex inference and prediction.” See Joshua Elliot, ARPA网站，9月访问。 2021年https://www.darpa.mil/program/artificial-social-intelligence-for-successful
“Artificial Social Intelligence for Successful Teams (ASIST),” DARPA website, accessed Sept. 20, 2021, -teams...
https://www.darpa.mil/program/artificial-social-intelligence-for-successful-teams.
Nagesh Jadhav and Rekha Sugandhi, “Survey on Human Behavior Recognition Using Affective Computing,” in
123 123Nagesh Jadhav和Rekha Sugandhi，"使用情感计算进行人类行为识别的调查"，2018年IEEE无线计算和网
2018 IEEE Global Conference on Wireless Computing and Networking (GCWCN), Lonavala, India, Nov. 23-24, 络全球会议（GCWCN），Lonavala，印度，11月11日。 23-24,2018,IEEE.
2018, IEEE.
124Samira Ranaei and Arho Suominen, “Using Machine Learning Approaches to Identify Emergence: Case of 124Samira Ranaei和Arho Suominen，"使用机器学习方法识别出现：车辆相关专利数据案例"，2017年波
Vehicle Related Patent Data,” in 2017 Portland International Conference on Management of Engineering and 特兰工程管理国际会议和
Technology (PICMET), Portland, OR, July 9-13, 2017, IEEE. 技术（PICMET），波特兰，OR，2017年7月9-13日，IEEE。
125James L. Regens, “Augmenting Human Cognition to Enhance Strategic, Operational, and Tactical Intelligence,” 125James L.Regens，"增强人类认知以增强战略，运营和战术情报"，《情报与国家安全34》，第5期（2019
Intelligence and National Security 34, no. 5 (2019). 年）。
38 38
scenario generation,126 adaptive and unpredictable opposing forces,127 and various forms of 场景生成，126适应性和不可预测的对立力量，127和各种形式的虚拟环境，128包括部分
virtual environments,128 including those generated in part using agent-based models.129 使用基于代理的模型生成的那些。 129
AI/ML and military wargaming are a particularly potent combination that can be applied in AI/ML和军事战争游戏是一个特别有效的组合，可以通过两种（部分重叠）方式应用：（1
two (partly overlapping) ways: (1) using traditional (non–AI/ML-enabled) wargames as ）使用传统的（非AI/ML启用的）战争游戏作为探索性教具，帮助确定AI可能对现有军事行
exploratory teaching aids to help identify the holistic effects that AI may have on existing 动概念（如CPLC中出现的概念）的整体影响，130和（2）通过将AI/ML技术直接嵌入战争
游戏来利用它们。在后一种情况下，AI/ML可用于开发"无所不知"的类似Alexa的前端界面
military operational concepts (such as those that appear in the CPLC),130 and (2) harnessing ，该界面有助于传统的wargaming实践131或作为天生的决策支持工具来"发现"替代行动方
AI/ML technologies by directly embedding them within wargames. In the latter case, AI/ML 案。 132最近的一项调查总结了最先进的AI/MLenabled自适应学习系统。 133
can be used to develop an “all-knowing” Alexa-like front-end interface that facilitates
conventional wargaming practices131 or as an innate decision-support tool to “discover”
alternative courses of action.132 A recent survey summarizes the state-of-the-art AI/ML-
enabled adaptive learning systems.133
Before making some general remarks about the gestalt of the CPLC-AI matrix in Figure 3, we 在对图3中CPLC-AI矩阵的格式塔进行一些一般性评论之前，我们注意到在该矩阵中作为列
note that implicit in all 11 of the domains that appear as columns in this matrix is AI’s inherent 出现的所有11个域中隐含的是AI固有的
126 Robert Sottilare, “A Hybrid Machine Learning Approach to Automated Scenario Generation (ASG) to Support 126Robert Sottilare在I3M Defense&Homeland Security Simulation中的"自动场景生成(ASG)的混合机器学习
Adaptive Instruction in Virtual Simulations and Games,” in I3M Defense & Homeland Security Simulation 方法以支持虚拟仿真和游戏中的自适应指令"
Workshop, Budapest, Hungary, 2018. 工作坊，匈牙利布达佩斯，2018年。
127Jeremy Ludwig and Bart Presnell, “Developing an Adaptive Opponent for Tactical Training,” in First 127Jeremy Ludwig和Bart Presnell，"为战术训练开发自适应对手"，在第一届自适应教学系统国际
International Conference on Adaptive Instructional Systems (AIS), Orlando, FL, July 26–31, 2019. 会议（AIS），佛罗里达州奥兰多，2019年7月26-31日。
Christina Cook, “Designing a Virtual Embedded Scenario-Based Military Simulation Training Program Using
128 128Christina Cook，"使用教育和设计教学策略设计基于虚拟嵌入式场景的军事模拟训练计划"（中佛罗里达
Educational and Design Instructional Strategies,” (Doctoral thesis, University of Central Florida, 2018). 大学博士论文，2018）。
129Wenhui Fan et al., “Multi-Agent Modeling and Simulation in the AI Age,” Tsinghua Science and Technology, 26, 129范文辉等。,"AI时代的多智能体建模与仿真",清华科技,26,no.5(Oct. 2021).
no. 5 (Oct. 2021).
130ED McGrady and Justin Peachy, eds., Representing Artificial Intelligence in Wargames, Connections Conference 130ED McGrady和Justin Peachy，eds。，在Wargames中代表人工智能，Connections Conference Working
Working Group 2 Proceedings, Dec. 2020, Group2Proceedings，Dec。 2020年，https://paxsims.files.wordpress.com/2020/12/connections_wg2_202
https://paxsims.files.wordpress.com/2020/12/connections_wg2_2020_final-4a.pdf. 0_final-4a.pdf...
131 Benjamin Jensen, Scott Cuomo, and Chris Whyte, “Wargaming with Athena: How to Make Militaries Smarter, 131Benjamin Jensen、Scott Cuomo和Chris Whyte《与雅典娜作战：如何利用人工智能使军队更智能、更快
Faster, and More Efficient with Artificial Intelligence,” War on the Rocks, June 5, 2018, 、更高效》《岩石上的战争》2018年6月5日https://warontherocks.com/2018/06/wargaming-with-athena
https://warontherocks.com/2018/06/wargaming-with-athena-how-to-make-militaries-smarter-faster-and- -how-to-make-militaries-smarter-faster-andmore-efficient-with-artificial-intelligence/.
more-efficient-with-artificial-intelligence/.
132AlphaZero has recently been applied to wargaming. See Glenn Moy and Slava Shekh, “The Application of 132AlphaZero最近被应用于wargaming。参见Glenn Moy和Slava Shekh,"The Application Of AlphaZero
AlphaZero to Wargaming,” in Advances in Artificial Intelligence, Proceedings of 32nd Australasian Joint to Wargaming,"In Advances in Artificial Intelligence,Proceedings of32nd Australasian Joint
Conference, Adelaide, SA, Australia, Dec. 2019. More focused discussions of how AI can be used for course-of- 会议,阿德莱德,SA,澳大利亚,12月. 2019. Peter J.Schwartz等人对如何将人工智能用于军事战争中的行动过程分
action analysis in military wargames are given by (1) Peter J. Schwartz et al., “AI-Enabled Wargaming in the 析进行了更集中的讨论。,"Ai-Enabled Wargaming in The Military Decision Making Process,"In Artificial Intellige
Military Decision Making Process,” In Artificial Intelligence and Machine Learning for Multi-Domain Operations nce and Machine Learning for Multi-Domain Operations Applications II,Proceedings Vol. 11413，2020和（2）
William DeBerry等人。,"战争商品课程
Applications II, Proceedings Vol. 11413, 2020, and (2) William DeBerry et al., “The Wargame Commodity Course of
Action Automated Analysis Method,” The Journal of Defense Modeling and Simulation: Applications, Methodology, 行动自动化分析方法，"国防建模与模拟杂志：应用，方法学,
Technology (July 19, 2021). Technology (July 19, 2021).
133Tumaini Kabudi, Ilias Pappas, and Dag Olsen, “AI-Enabled Adaptive Learning Systems: A Systematic Mapping of 图马尼*卡布迪、伊利亚斯*帕帕斯和达格*奥尔森，"人工智能支持的自适应学习系统：文献的系统映射"，计算
the Literature,” Computers and Education: Artificial Intelligence 2 (2021). 机与教育：人工智能2（2021年）。
39 39
ability to discover novelty in unimaginably vast “possibility spaces” that no human is 在难以想象的巨大"可能性空间"中发现新奇的能力，没有人能够快速或有效地解析。 134
capable of parsing as quickly or efficiently.134
Commentary about the CPLC-AI matrix in Figure 3 关于图3中CPLC-AI矩阵的评论

What does the CPLC-AI matrix in Figure 3 actually show us? The most obvious takeaway is that, 图3中的CPLC-AI矩阵实际上向我们展示了什么？最明显的外卖是，无一例外，26个CPLC
without exception, each of the 26 CPLC elements has significantly more than a single AI/ML 元素中的每一个都具有显着多于与其相关的单个AI/ML应用：最小值，最大值和平均值分别
application associated with it: the minimum, maximum, and average values are 2, 19, and 为2,19和7.3。同样，给定AI/ML应用与之相关的CPLC元件的最小、最大和平均数分别为1
、14和5.6。尽管可能会质疑任何特定条目的存在（或不存在），但无限丰富的潜在关联网
7.3, respectively. Similarly, the minimum, maximum, and average number of CPLC elements 络证明了AI/ML技术的巨大潜在普遍适用性。
with which a given AI/ML application is associated are 1, 14, and 5.6, respectively. Although
the presence (or absence) of any given entry may be questioned, the unassailably rich network
of potential associations testifies to the enormous latent general applicability of AI/ML
technologies.
Remember that AI/ML-enabled technologies that support each of the 11 domains that appear 请记住，支持图3中出现的11个领域中的每一个领域的AI/ML技术已经存在（尽管效果不同
in Figure 3 already exist (albeit to varying degrees of efficacy and subject to the vagaries of ，并且受到正在进行的基础研究的变幻莫测），这意味着可以利用广泛的AI/ML工具库来减
ongoing basic research), which means that a broad arsenal of AI/ML-enabled tools can be 轻平民伤害。可以肯定的是，魔鬼在细节中：一般平民伤害问题没有灵丹妙药"解决方案"
（无论是否启用AI/ML;下一节将讨论一些具体选项）。但仅仅是图3包含了如此多的合理分
harnessed to mitigate civilian harm. To be sure, the devil is in the details: there are no panacea 配的"uma"条目，将CPLC的元素与潜在的AI/ML应用相关联，这一事实就强烈地表明，可以
“solutions” to the general civilian harm problem (whether AI/ML enabled or not; some specific 立即完成大量尚未开发的事物库，以增强CPLC的所有组成部分。
options are discussed in the next section). But the mere fact that Figure 3 contains so many
plausibly assigned “◼” entries associating elements of the CPLC with potential AI/ML
applications strongly suggests that a vast heretofore largely untapped reservoir of things can
immediately be done to enhance essentially all components of the CPLC.
Table 1 shows the number of links in the CPLC-AI matrix within each of the main (i.e., top-level) 表1显示了每个主要（即顶级）内CPLC-AI矩阵中的链接数
CPLC and AI/ML application elements. CPLC和AI/ML应用元件。
Table 1. Number of “ ” entries in each of the top-level categories in the CPLC-AI matrix in Table 1. CPLC-AI矩阵中每个顶级类别的""条目数
Figure 3 Figure 3
Artificial Intelligence (AI) / Machine Learning (ML) Applications 人工智能(AI)/机器学习(ML)应用
USVs Cyber CV ET KR/R NLP P&S DM Tgt TM Trn Total USVs Cyber CV ET KR/R NLP P&S DM Tgt TM Trn Total
M&M 1 1 0 1 2 3 4 4 0 0 4 20 M&M 1 1 0 1 2 3 4 4 0 0 4 20
Civilian Civilian
Protection Planning 1 4 5 2 8 1 6 6 2 6 9 50 Protection Planning 1 4 5 2 8 1 6 6 2 6 9 50
Life 35 Life 执行董事 35
Op Exec 2 1 3 5 3 0 5 3 5 6 2 2 1 3 5 3 0 5 3 5 6 2
134See Joel Lehman et al., “The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the 134见Joel Lehman等人。，"数字进化令人惊讶的创造力：进化计算和人造生命研究社区的轶事集合，"1
Evolutionary Computation and Artificial Life Research Communities,” Nov. 21, 2019, arXiv:1803.03453,; and 1月11日。 21，2019，arXiv：1803.03453，;和
Giorgio Franceshelli and Mirco Mosolesi, “Creativity and Machine Learning: A Survey,” Apr. 20, 2021, Giorgio Franceshelli和Mirco Mosolesi，"创造力和机器学习：一项调查"，Apr。 20，2021，arXiv
arXiv:2104.02726v2. ：2104.02726v2。
40 40
Artificial Intelligence (AI) / Machine Learning (ML) Applications 人工智能(AI)/机器学习(ML)应用
Cycle Assess 2 2 2 6 5 4 0 2 2 3 3 31 Cycle Assess 2 2 2 6 5 4 0 2 2 3 3 31

(CPLC) (CPLC)
Response 1 9 0 2 2 11 1 1 1 2 1 32 Response 1 9 0 2 2 11 1 1 1 2 1 32
Lrn&Adapt 0 4 0 0 4 0 2 3 0 2 6 21 Lrn&Adapt 0 4 0 0 4 0 2 3 0 2 6 21
Total 7 21 10 16 24 19 18 19 10 19 25 Total 7 21 10 16 24 19 18 19 10 19 25
The plethora of potential applications ought not surprise us because the problem of civilian 过多的潜在应用不应该让我们感到惊讶，因为平民伤害问题可能被视为与更大的军事行动
harm may be viewed as a microcosm of actions, behaviors, and policies associated with the 空间相关的行动，行为和政策的缩影。而且，正如许多研究（和现实世界的行动）所表明
much larger military operational space. And, as numerous studies (and real-world operations) 的那样，几乎没有任何军事相关的活动领域不能应用AI/MLenabled技术，即使只是原则上
。 135
demonstrate, there is virtually no militarily relevant domain of activity to which AI/ML-
enabled technologies cannot be applied, even if only in principle.135
As argued earlier, the CPLC-AI matrix in Figure 3 (and its derivative tallies in Table 1) ought 如前所述，图3中的CPLC-AI矩阵（及其在表1中的衍生物计数）除了其总体可信度和其各
not be imbued with any deeper meaning apart from its overall general credibility and the 个条目的知情合理性之外，不应具有任何更深层次的含义。然而，我们指出了一些可能不
informed plausibility of its individual entries. However, we point out a few salient features that 会立即明显的显着特征:
may not be immediately obvious:
• A few entries are effectively “one offs,” meaning that the range of their associability • 一些条目实际上是"一个取舍"，这意味着它们的关联性（或适用性）的范围本身仅
(or applicability) is natively limited to no more than a few cases. For example, “damage 限于不超过少数情况。例如，图3第一列中通过自主UASs/USVs进行的"损害评估"
assessment” via Autonomous UASs/USVs in the first column of Figure 3 is an innately 是一个天生集中的应用程序，仅限于支持CPLC评估元素的"评估战斗损害"次要组成
部分，而CPLC规划元素下第八行中的"为CHM分配能力"则需要一组固有的小型相关
focused application limited to supporting the “assess battle damage” secondary AI/ML能力。
component of CPLC’s Assessment element, and “allocate capabilities for CHM” in the
eighth row under the CPLC’s Planning element entails an inherently small set of
relevant AI/ML-enabled capabilities.
• Some CPLC elements and AI/ML applications are only marginally associable. • 一些CPLC元件和AI/ML应用仅略微关联。示例包括表1中零条目的那些，例如学
Examples include those with zero entries in Table 1, such as the intersections of 习和适应以及自主UASs/USVs，响应和CV以及操作的交叉点
Learning and Adapting and Autonomous UASs/USVs, Response and CV, and Operational
Execution and NLP. 执行和NLP。
135 An extensive taxonomy of AI applications to military operations (woven partly around elements of the OODA 135.人工智能应用于军事行动的广泛分类（部分围绕OODA循环的元素编织：观察、定向、决定和行动）出现
loop: observe, orient, decide, and act) appears in Appendix I, “Mindmap of Possible Military Applications of AI,” in 在附录I"人工智能可能的军事应用的思维导图"中，在Ilachinski，人工智能：新兴的主题、问题和叙述中。有关
Ilachinski, Artificial Intelligence: Emerging Themes, Issues, and Narratives. For a basic taxonomy, see Stoney Trent 基本分类法，请参阅Stoney Trent和Scott Lathrop，"军事领导人人工智能入门"，Small Wars Journal（Aug。 2
018).
and Scott Lathrop, “A Primer on Artificial Intelligence for Military Leaders,” Small Wars Journal (Aug. 2018).
41 41
• Certain specific AI/ML capabilities that fall under a given broad application area • 属于给定广泛应用领域的某些特定AI/ML能力可应用于多于一个CPLC元件，包括表
may be applied to more than one CPLC element, including those that do not 面上不属于相同区域的那些。例如，虽然培训可能不会直观地预期除了例如规划
ostensibly belong to the same area. For example, although Training may not 或学习和适应之外的CPLC元素发挥作用，但基本的基础方法适用于CPLC的所有部
分。
intuitively be expected to play a role for CPLC elements other than, for example,
Planning or Learning and Adapting, the basic underlying methodologies are applicable
to all parts of the CPLC.
• The main CPLC elements with the largest average number of associated AI/ML • 相关AI/ML应用程序平均数量最大的主要CPLC元素是操作执行和规划。它们的平
applications are Operational Execution and Planning. They have averages of 43/4 均值分别为43/4x10.8和54/6x9.0(见表1)。除了缺乏机器人技术对规划的任何明显
适用性之外，这两个要素最共同的是需要在潜在的巨大抽象"可能性空间中进行裁决
 10.8 and 54/6  9.0, respectively (see Table 1). Except for the lack of any obvious
。"这是ML特别擅长的，以及结合多种同时功能，如因果推理，多维风险评估和AI/
applicability of robotic technologies to planning, what these two elements have most ML启用的wargaming。
in common is a need to adjudicate among a potentially massive abstract “space of
possibilities.” This is something ML is particularly adept at, as well as combining
multiple simultaneous capabilities, such as causal reasoning, multi-dimensional risk
assessment, and AI/ML-enabled wargaming.
• The secondary-level CPLC element with the largest number of associated AI/ML • 关联AI/ML应用程序数量最多的二级CPLC元素（即利用AI/ML启用技术的最大内在
applications (i.e., the element with the greatest innate potential for leveraging 潜力的元素）是正在规划中的"平民生活模式"。它的19个关联包括来自每个顶级应
AI/ML-enabled technology) is “pattern of life for civilians” under Planning. Its 19 用程序域的至少一个AI/ML应用程序。关联数量次多的要素是行动执行下的"战术耐
心"（有18个关联），其次是任务和任务下的"塑造特派团以减轻平民的风险"；"塑
associations include at least one AI/ML application from each of the top-level 造行动以减轻风险"和"制定战术替代方案"，都在规划中；以及评估下的"分析平民伤
application domains. The element with the next largest number of associations is 害数据"（各有11个相关的AI/ML应用程序）。启发式地，我们期望这种多方面的"
“tactical patience” under Operational Execution (with 18 associations), followed by 问题"需要丰富和互补的方法和技术（例如，传感器，行为识别，推理引擎）组装。
“shape the mission to mitigate risks to civilians” under Mission and Mandate; “shape
operations to mitigate risks” and “develop tactical alternatives,” both under Planning;
and “analyzing civilian harm data” under Assessment (all with 11 associated AI/ML
applications each). Heuristically, we expect such multifaceted “problems” to require a
rich and complementary assembly of methods and technologies (e.g., sensors, behavior
recognition, inference engines).
• The AI/ML applications domains associated with the largest number of CPLC • 与最大数量的CPLC元素相关的AI/ML应用领域是Training（26）和KR＆R（24）。
elements are Training (26) and KR&R (24). Cyberspace is a close third with 23 网络空间近三分之一，有23个协会。这三个领域的基础方法分别属于更广泛和研
associations. The methods underlying these three areas fall under much broader and 究良好的基于复杂系统的multiagent模型，模拟和知识图（编纂整个语义和因果本
体）的类别。因此，它们也是最普遍地与CPLC元件的色域相关联的，这并不奇怪
well-studied classes of complex systems–based multiagent models, simulations, and 。
knowledge graphs (that codify entire semantic and causal ontologies), respectively. It
is thus not surprising that they are also the most generally associable with the gamut
of CPLC elements.
• The specific AI/ML applications with the largest number of associated CPLC • 相关CPLC元素数量最多的特定AI/ML应用是训练下的"基于代理的模型/战争游戏/虚
elements are “agent-based models/wargames/virtual environments” under 拟环境"（15），网络空间下的"社交网络/众包映射"（12）和P＆S下的"不确定性
Training (15), “social networks/crowdsourced mapping” under Cyberspace 下的规
(12), and “planning under uncertainty” under P&S (11), although “causal
reasoning” (under KR&R) and “language analytics/summarization” (under NLP), with
42 42
10 associations each, are not far behind. These are all methodologically and 各10个协会，也不甘落后。这些都是方法和技术成熟的应用程序，随时可以利用
technologically mature applications ready to be harnessed. 。
What Figure 3 does not show, at least explicitly, is that multiple simultaneous algorithms, 图3没有显示的是，至少明确地说，多个同时进行的算法、应用程序和技术可以协同组合，
applications, and technologies can be combined synergistically to develop powerful standalone 以开发功能强大的独立多用途AI/ML系统。
multipurpose AI/ML-enabled systems.
There is perhaps no better example of this than GAIA (Generating Alternative Interpretations 也许没有比GAIA（生成用于分析的替代解释）更好的例子了，该解释部分由DARPA资助，
for Analysis), funded in part by DARPA and developed by the US Army Research Laboratory, 由美国陆军研究实验室，伊利诺伊大学香槟分校和哥伦比亚大学开发。盖亚在2020年7月
the University of Illinois at Urbana‐Champaign, and Columbia University. GAIA won the “Best 的计算语言学协会第58届年会上获得了"最佳演示"奖。它被开发者描述为"细粒度多媒体知
识提取系统"，是"第一个全面的开源多媒体知识提取系统，它将来自各种来源和语言的大量
Demo” award at the 58th Annual Meeting of the Association for Computational Linguistics in 非结构化异构多媒体数据作为输入，并创建一个连贯的、结构化的知识库、索引实体、关系
July 2020.136 It is described by its developers as a “fine-grained multimedia knowledge 和事件，遵循丰富的细粒度本体。"137GAIA结合了多种同步ML方法，包括文本和视觉知识
extraction system” and is the “first comprehensive, open source multimedia knowledge 提取和关联，公众人物的目标检测和识别，跨媒体知识融合（文本，图像，视频，语音和光
extraction system that takes a massive stream of unstructured, heterogeneous multimedia data 学字符识别）以及行它提供了对正在进行的事件的实时跟踪，问题的上下文和动态相关的
from various sources and languages as input, and creates a coherent, structured knowledge 警报，并包括一个名义上的能力"预测"对环境或操作的变化，以及预测可能与正在进行的事
件有关的其他主题和数据。 138
base, indexing entities, relations, and events, following a rich, fine-grained ontology.”137 GAIA
combines multiple simultaneous ML methods, including text and visual knowledge extraction
and correlation, object detection and recognition of public figures, cross-media knowledge
fusion (text, images, videos, speech, and optical character recognition), and behavioral and
evidential pattern recognition. It provides real‐time tracking of ongoing events, issues
contextually and dynamically relevant alerts, and includes a nominal ability to “predict”
changes to an environment or operations as well anticipate other topics and data that might be
related to ongoing incidents.138
The Command and Control Incident Management Emergency Response Application 指挥和控制事件管理应急响应应用程序（C2imera）系统是一个大规模军事级工具（尽管是
(C2IMERA) system is an example of a large-scale military-grade tool (albeit one that uses AI 一个以基本方式使用AI和ML的工具）的例子，它已经部署并提供规划、部队生成、应急管
and ML in a rudimentary fashion) that is already deployed and provides planning, force 理、实时SA和指挥和控制（C2）监测和执行功能。 139它是由空军生命周期管理中心开发
的
generation, emergency management, real-time SA, and command and control (C2) monitoring
and execution functions.139 It was developed by the Air Force Life Cycle Management Center’s
136Manling Li et al., “GAIA: A Fine-Grained Multimedia Knowledge Extraction System,” in Proceedings of the 58th 136李曼玲等人。,"GAIA:A Fine-Grained Multimedia Knowledge Extraction System,"In Proceedings of The58
Annual Meeting of the Association for Computational Linguistics: System Demonstrations, July 2020, th Annual Meeting Of The Association For Computational Linguistics:System Demonstrations,July2020,https:/
https://aclanthology.org/2020.acl-demos.11v2.pdf. /aclanthology.org/2020.acl-demos.11v2.pdf...
137 Li et al., “GAIA.” p. 1. 137Li等人。，"盖亚。"第1页。
138GAIA: A Fine-Grained Multimedia Knowledge Extraction System (Blender Lab, UIUC), Video, 138GAIA:A Fine-Grained Multimedia Knowledge Extraction System(Blender Lab,UIUC),Vi
http://blender.cs.illinois.edu/software/gaia-ie/gaia.mp4. deo,http://blender.cs.illinois.edu/software/gaia-ie/gaia.mp4...
139US Air Force, C2IMERA Overview (Kessel Run, Mar. 26, 2020), Video, 139美国空军，C2IMERA概述（Kessel Run，Mar。 26，2020），视频，https://www.youtube.com/watch
https://www.youtube.com/watch?v=8meuTt1RQbE. C2IMERA is mandated for use across all Air Combat ?v=8meuTt1RQbE。C2IMERA被授权用于所有空战指挥装置，目前在美国大陆的40多个美国空军基地使用
Command installations and is currently used at more than 40 US Air Force bases across the continental United
States and Europe. 国家和欧洲。
43 43
Detachment 12 (also known as “Kessel Run”), a software development and acquisitions unit.140 12支队（也称为"Kessel Run"），一个软件开发和收购单位。 140C2IMERA结合了两个核心
C2IMERA combines two core C2 capabilities, real-time SA and coordination, by making all C2功能，实时SA和协调，通过在一个集中位置提供有关特定安装，环境，资产和人员的所
pertinent data about specific installations, environments, assets, and personnel accessible in 有相关数据。在2019中，佐治亚州穆迪空军基地的C2专家使用它来跟踪飓风多利安，帮助
为即将到来的风暴做好准备，并确定飓风路径之外的飞离选项。 141最近在阿富汗C2IMER
one centralized location. In 2019, it was used by C2 specialists at Moody Air Force Base, A被用作减轻平民伤害的工具以支持最近的非战斗人员撤离行动。 142随着从阿富汗撤离的
Georgia, to track Hurricane Dorian, helping prepare the base for the approaching storm and 发生，C2IMERA被用来跟上快速变化的环境，并确保超过124,000美国文职和军事人员，盟
identifying fly-away options outside the path of the hurricane.141 More recently, in Afghanistan, 友和合作伙伴以及来自喀布尔的阿富汗人的安空中和空间联合行动中心还使用C2IMERA接
C2IMERA was used as a civilian harm-mitigation tool to support the recent noncombatant 收事件和响应的自动警报，这提供了近乎实时的额外协调水平。 143虽然C2IMERA目前仅
evacuation operation.142 As the evacuations from Afghanistan took place, C2IMERA was used 最低限度地利用AI或ML，但这一已经验证和部署的系统的几乎所有元素都只能通过集成各
种支持AI/ML的技术来增强。
to keep abreast of a rapidly changing environment and ensure the safe transit of more than
124,000 US civilian and military personnel, allies and partners, and Afghans from Kabul. The
Combined Air and Space Operations Center also used C2IMERA to receive automated alerts of
the incident and response, which provided an additional level of coordination in near real
time.143 Although C2IMERA currently leverages AI or ML only minimally, virtually all elements
of this already proven and deployed system can only be enhanced by integrating various
AI/ML-enabled technologies.
Now that we have gotten a bird’s eye view of the landscape of CPLC and AI/ML (and discussed 现在我们已经鸟瞰了CPLC和AI/ML的景观（并讨论了一系列可能的关联），接下来我们将
a litany of possible associations), we next examine some specific ways in which various 研究一些具体的方式，使各种潜在的平民伤害缓解措施可以从AI/ML使能的技术中受益。在
potential mitigations for civilian harm may benefit from AI/ML-enabled technologies. In the 前一章中我们讨论了造成平民伤害的不同机制以及为减少对平民的风险而需要解决的一些具
体问题。我们再次讨论了其中的一些风险，但以不同的方式，强调了人工智能/ML技术将运
previous chapter, we discussed the different mechanisms that cause civilian harm and some 行的环境的局限性。
specific problems that need to be addressed to reduce risks to civilians. We discuss some of
these risks again but in a different way, highlighting the limitations of the environment in
which AI/ML-enabled technologies will be operating.
Reducing but not eliminating risks to civilians 减少但不消除对平民的风险

Figure 4 shows a notional schematic illustrating key elements of the environment many AI 图4显示了一个名义上的原理图，说明了许多寻求减轻平民伤害的人工智能应用程序将要处
applications seeking to mitigate civilian harm will be dealing with. It highlights the critical 理的环境的关键要素。它突出了操作区域中的关键代理和特征，包括目标区域（以紫色表
agents and features in an operating area, including the targeting area (indicated in purple), 示）,
140 US Air Force, Kessel Run homepage, https://kesselrun.af.mil/. 140美国空军,凯塞尔运行主页,https://kesselrun.af.mil/。

141 “Adaptable Command and Control System Allowed Georgia Air Force Base to Track and Prepare for Hurricane 141"适应性指挥和控制系统使格鲁吉亚空军基地能够跟踪和准备飓风多利安，"Leidos，10月14日。 2019年1
Dorian,” Leidos, Oct. 1, 2019, https://www.leidos.com/insights/adaptable-command-and-control-system- 月1日https://www.leidos.com/insights/adaptable-command-and-control-systemallowed-georgia-air-force
allowed-georgia-air-force-base-track-and-prepare. -base-track-and-prepare.
142Richard Blumenstein, “Kessel Run’s C2IMERA Used During Afghan Evacuation,” Wright-Patterson AFB, Sept. 23, 142Richard Blumenstein，"Kessel Run的C2imera在阿富汗撤离期间使用，"Wright-Patterson AFB，9月。 202
2021, https://www.wpafb.af.mil/News/Article-Display/Article/2786182/kessel-runs-c2imera-used-during- 1年https://www.wpafb.af.mil/News/Article-Display/Article/2786182/kessel-runs-c2imera-used-duringafghan
afghan-evacuation/. -evacuation/.
143Mark Pomerleau, “Air Force Software Tool Helped Coordinate Afghanistan Evacuation of Civilians,” C4ISRNet, 143Mark Pomerleau，"空军软件工具帮助协调阿富汗平民撤离"，C4ISRNet，9月9日。 2021年https://www.
Sept. 23, 2021, https://www.c4isrnet.com/battlefield-tech/c2-comms/2021/09/23/air-force-software-tool- c4isrnet.com/battlefield-tech/c2-comms/2021/09/23/air-force-software-toolhelped-coordinate-afghanis
helped-coordinate-afghanistan-evacuation-of-civilians/. tan-evacuation-of-civilians/.
44 44
combatants (red), friendly military forces (blue), civilians (dark green), and surrounding 战斗人员（红色），友军（蓝色），平民（深绿色）和周围
area, including buildings (light green rectangles, with a superimposed red “✚” denoting a 区域，包括建筑物（浅绿色矩形，叠加红色"✚"表示医院）和基础设施（由建筑物之间的链
hospital) and infrastructure (designated by links between buildings). Leaving aside for the 接指定）。暂时撇开某些不可减少的现实世界的不确定性，例如战斗人员和平民的确切人
数和地点（详见下文的例子）不谈，这些是任何人工智能/机器学习技术必须能够感知、理
moment certain irreducible real-world uncertainties, such as the exact numbers and locations 解、推断并在减轻平民伤害的情况下作出决定的基本要素。接下来的几个数字表明，即使
of combatants and civilians (detailed in the examples that follow), these are the basic elements 是一个很少或没有不确定性的完美世界，在"敌对"和"非目标"之间有明确的界限，并且其中
that any AI/ML-enabled technology must be able to sense, understand, draw inferences from, 排除任何合理的附带损害可能性的目标区域（以及伴随的武器爆炸区）都很容易识别，对平
and otherwise make decisions about in the context of mitigating civilian harm. The next few 民也会产生非零风险。 144
figures show that even a perfect world⎯one with few or no uncertainties, with clear
demarcations between “hostile” and “nonhostile,” and in which targeting areas (and
concomitant weapon blast zones) that preclude any reasonable likelihood of collateral damage
are all easily identifiable⎯will have a non-zero risk to civilians.144
Figure 4. Notional schematic illustrating key elements of the civilian harm AI problem 图4。说明平民伤害AI问题关键要素的名义示意图
Why is a zero-risk scenario unattainable? Consider Figure 5, which adds a few changes to 为什么零风险情景无法实现？考虑图5，其中增加了一些更改
Figure 4 to reflect the nature of real-world ambiguities: 图4反映现实世界模糊的性质:
1. Buildings may exist that do not appear in out-of-date or incomplete targeting datasets. 1. 可能存在未出现在过时或不完整的目标数据集中的建筑物。
144This is an important point because some commanders in the field have stressed a goal of no CIVCAS. Over the 144这是一个重要的观点因为实地的一些指挥官强调了一个没有民警的目标。从长远来看，即使利用人工智
long term, even with the use of AI to mitigate civilian harm more effectively, this laudable goal is not possible. 能更有效地减轻平民伤害，这一值得称赞的目标也是不可能的。但是，我们仍然可以使用技术来降低这些风
However, we can still use technology to reduce these risks. 险。
45 45
2. The “targeting status” of other known buildings may be changed (e.g., a warehouse 2. 其他已知建筑物的"目标状态"可能会改变（例如，仓库可能已转变为宗教中心）。
may have been transformed into a religious center).
3. Civilian transients may have entered an otherwise properly designated targeting area 3. 在发出"攻击"命令之前，平民瞬变可能已经进入了其他适当指定的目标区域。
immediately before an “attack” order was issued.
4. An old, weathered (maybe already combat-torn) protected symbol designating, for 4. 一个古老的，风化的（可能已经被战斗撕裂的）保护符号，例如医院，可能无法识
example, a hospital, may go unrecognized and be incorrectly targeted. 别并且被错误地定位。
Figure 5. Schematic adding some inherent uncertainty to the notionally “perfect” setting in 图5。原理图在概念上"完美"设置中添加了一些固有的不确定性
Figure 4 Figure 4
Figure 6 adds four more layers of ambiguity and uncertainty: 图6增加了四层模糊性和不确定性:
5. The targeting area may be moved to another location (do military units have sufficient 5. 瞄准区域可能会移动到另一个位置（军事单位是否有足够的和最新的SA？).
and up-to-the-minute SA?).
6. Correctly identified hostiles that were previously outside the original targeting area 6. 以前在原瞄准区以外的正确识别的敌人可能会被错误地误认为无意中误入新瞄准区
may be erroneously misidentified with civilians that inadvertently stray into the new 的平民。
targeting area.
7. Hostiles that were previously in the newly designated targeting area maneuver 7. 以前在新指定的目标区域的敌人在它外面机动。
outside of it.
8. Otherwise-legitimate military targets within the new target area are parts of a larger 8. 否则，新目标区内的合法军事目标是更大基础设施的一部分，其中包括如果受到攻
infrastructure that includes elements which, if attacked, would harm civilians. 击会伤害平民的元素。示例包括明确的两个示例，例如电网（可能更容易
Examples include both those that are explicit, like power grids (which may be easier
46 46
to detect visually and incorporate into targeting strategies), and those that are implicit, 以视觉检测并纳入目标策略），以及那些隐含的，如下水道系统（可能单独从视觉
like a sewer system (which may be hidden from visual inspection alone). 检查中隐藏）。
Figure 6. Schematic adding a second layer of ambiguity or uncertainty to Figure 4 图6。向图4添加第二层模糊性或不确定性的示意图
Although far from complete, Figure 5 and Figure 6 highlight some of the ways in which harm 图5和图6虽然还远远没有完成，但突出了对平民造成伤害（包括平民伤害）的一些方式。
to civilians (including civilian harm) can arise. These are consistent with the mechanisms 这些机制与图1中提供的机制一致（但使用真实世界的示例可以为我们提供有关这些机制的
provided in Figure 1 (but using real-world examples gives us more detail on those 更多详细信息），并添加其他注意事项，例如保护关键基础设施。
mechanisms), with the addition of other considerations such as the protection of critical
infrastructure.
The potential for harm to civilians and critical infrastructure can be mitigated in practice by 在实践中，可以通过以下方式减轻对平民和关键基础设施造成伤害的可能性:
the following:
• Continually monitoring all militarily significant and civilian elements within an • 以最大可能的保真度持续监测行动区内的所有军事重要人员和文职人员
operating area with the greatest possible fidelity
• Continually assessing, and drawing predictive actionable inferences from, the overall • 持续评估操作区域内的整体活动（即其"生活模式"），并从中得出可预测的可操作
activity within the operating area (i.e., its “pattern of life”) 推论
• Recognizing, on a fundamental “system of systems” level of analysis, that all elements • 认识到在基本的"系统系统"分析水平上，军事行动区域内的所有要素都是有机和动
within a military operational area are organically and dynamically entwined 态地交织在一起的（以及随之而来的对整体瞄准和损害评估的影响）
(along with attendant implications for holistic targeting and damage assessment)
47 47
We will now leverage the general relationship between the CPLC and AI/ML applications 我们现在将利用上一节中介绍的CPLC和AI/ML应用程序之间的一般关系来展示如何使用特
introduced in the last section to show how specific AI/ML-enabled technologies can be brought 定的AI/ML技术来解决这个问题。
to bear on this problem.
CIVCAS Cause-Mitigations-AI-applications CIVCAS Cause-Mitigations-AI-applications

(CIVCAS-AI) matrix (CIVCAS-AI) matrix
Figure 7 shows the CIVCAS Cause-Mitigations-AI-applications (CIVCAS-AI) matrix. The 图7显示了CIVCAS Cause-Mitigations-AI-applications(CIVCAS-AI)矩阵。该图实际上包含
figure actually contains two linked matrices: one that associates potential civilian harm 两个相互关联的矩阵：一个将潜在的平民伤害减轻与四个主要的平民伤害原因中的每一个联
系起来（其条目在左边以红色突出显示），另一个将潜在的平民伤害减轻与特定的AI/ML应
mitigations with each of four leading causes of civilian harm (and whose entries are highlighted
用（图3中出现的相同的应用）联系起来（其条目在右边以蓝色突出显示）。第二个矩阵中
in red on the left) and another that associates potential civilian harm mitigations with specific 括号内的数字（以绿色和蓝色突出显示）分别表示出现在相应行和列中的"◼"条目的数量。
AI/ML applications (the same ones that appear in Figure 3) civilian harm(and whose entries 带有符号的条目"
are highlighted in blue on the right). The numbers in parentheses in the second matrix
(highlighted in green and blue), denote the number of “◼” entries that appear in the
corresponding row and column, respectively. Entries with the symbol “ ” identify the most "为给定的民用伤害缓解确定最直观的可harnessable AI/ML应用程序。
intuitively harnessable AI/ML applications for a given civilian harm mitigation.
Figure 7. CIVCAS Cause-Mitigations-AI-applications (CIVCAS-AI) Matrix 图7。 CIVCAS原因-缓解-AI-应用(CIVCAS-AI)矩阵
48 48
The potential civilian harm mitigations listed in the middle column of Figure 7 mainly address 图7中间一栏中列出的潜在平民危害缓解措施主要涉及规划和运营执行阶段（仅仅是因为我
the Planning and Operational Execution stages (simply because we expect AI/ML-enabled 们期望启用AI/ML的应用程序在这两个阶段内具有最大的可缓解措施要么颁布或扩展特定的
applications to have the maximum leverageable impact within these two phases). Mitigations CPLC要素（例如，规划下的"生活模式"和操作执行下的"战术耐心"，以便相关的应用程序直
接反映图3中出现的应用程序），要么隐含地结合一个或多个要素（例如，"适应性武器化"
either enact or extend a specific CPLC element (e.g., “pattern of life” under Planning and ，同时处理规划下的"塑造行动以减轻风险"和操作执行下的"武器化以减轻风险"）。
“tactical patience” under Operational Execution so that associated applications directly mirror
those that appear in Figure 3) or implicitly combine one or more elements (e.g., “adaptive
weaponeering,” which simultaneously addresses “shape operations to mitigate risks” under
Planning and “weaponeering to mitigate risks” under Operational Execution).
The right-most matrix in Figure 7 shows that each of the nine potential civilian harm 图7中最右边的矩阵显示，九种潜在的民用伤害缓解措施中的每一种都可以从多种AI/ML应
mitigations stands to benefit from multiple kinds of AI/ML applications. Indeed, it was 用中受益。事实上，正是着眼于实现这种潜在的协同作用，我们才有动力制定这份清单。
我们依次讨论九项缓解措施中的每一项。
with one eye focused on achieving such a potential synergy that we were motivated to
construct this list. We discuss each of the nine mitigation measures in turn.
Robust civilian pattern of life. Tools that can help human decision-makers better understand 稳健的平民生活模式。可以帮助人类决策者更好地理解平民生活模式的工具同样来自多个
civilian patterns of life similarly derive from multiple overlapping AI/ML technology domains, 重叠的AI/ML技术领域，从AI/ML启用的通用预测智能工具开始。其他应用包括使用传感器
starting with AI/ML-enabled tools for general predictive intelligence. Other applications 群、对象识别和分类、行为模式分析和语言分析来开发健壮的数据集以供进一步分析；对这
些数据集应用众包映射、因果推理和潜在模式识别技术来推断（或"发现"）其他潜在行为；
include using sensor swarms, object recognition and classification, behavior pattern analysis, 以及使用ABM技术开发数据化模拟，以帮助预测目标需求并制定一般的民用基于风险的政
and language analytics to develop robust datasets for further analysis; applying crowdsourced 策选择。
mapping, causal reasoning, and latent pattern recognition techniques to those datasets to infer
(or “discover”) otherwise latent behaviors; and using ABM techniques to develop data-
informed simulations to help anticipate target requirements and develop general civilian risk–
based policy options.
Holistic collateral damage estimate, target development, and adaptive weaponeering. 整体附带损害估计,目标发展和适应性武器装备. 虽然这三种潜在的平民伤害减轻措施都有
Although each of these three potential civilian harm mitigations has unique requirements (as 独特的要求（它们的相关应用程序集略有不同就证明了这一点），但它们都在很大程度上取
evidenced by their slightly differing sets of associated applications), they all depend critically 决于它们能够利用一个关键能力的程度，即感知、整合和预测全系统动态要素的能力。当
然，至于其他民用伤害缓解措施，这些缓解工具的效用取决于有其他AI/ML应用程序来支持
on the degree to which they can leverage one key capability, namely the capacity to perceive, 它们（例如，使用复杂的网络分析和模式识别工具来研究和发现通过系统的材料和信息流的
integrate, and anticipate systemwide dynamical elements. Of course, as for other civilian harm 连接模式）。重要的是，这三个应用程序中的每一个都将通过更好地告知目标确定过程中
mitigations, the utility of these mitigation tools depends on having other AI/ML applications to 平民和相关基础设施面临的实际风险程度来提高决策的质量。
support them (e.g., the use of complex network analysis and pattern recognition tools to
study⎯and discover⎯connectivity patterns of materiel and information flow through a
system). Of importance, each of these three applications would enhance the quality of decision-
making by better informing the actual magnitude of risks to civilians and associated
infrastructure during the targeting process.
Tactical patience. Tactical patience is a term originally coined in 2009 during the Afghanistan 战术耐心。战术耐心是最初在阿富汗战役期间在2009中创造的一个术语。在该操作中，
campaign. A key element of command guidance in that operation, it involves three of the four 命令指导的一个关键元素，它涉及OODA循环的四个元素中的三个（即观察，定向和决定）
elements of the OODA loop (namely, observe, orient, and decide). Exercising tactical patience 。行使战术耐心意味着部队需要额外的时间来获取（和考虑）更准确的信息以做出更好的
决策。这样做也提供了重新审视的机会
means that forces take additional time as available to acquire (and deliberate about) more
accurate information to make better decisions. Doing so also affords an opportunity to revisit
49 49
earlier or existing assumptions and incomplete or imprecise data, consider alternative courses 较早或现有的假设以及不完整或不精确的数据考虑备选行动方案并探讨减轻无处不在的"战
of action, and explore options to mitigate ubiquitous “fog of war” uncertainties, particularly 争迷雾"不确定性的备选方案特别是那些可能无意中导
those that may inadvertently lead to civilian harm.
Although the class of AI/ML-enabled tools designed to facilitate “multi-dimensional risk 虽然旨在促进决策下的"多维风险评估"的ai/ML启用工具的类别可能有望提供最明显的即时
assessment” under Decision-Making may be expected to provide the most obvious immediate 收益（如符号所强调的"
benefit (as highlighted by the symbol “ ” in Figure 7), many other technologies can also be "在图7中），许多其他技术也可以利用。一些可能性包括（1）利用智能传感器网格，社交
harnessed. Some possibilities include (1) leveraging smart sensor grids, predictive sentiment 网络的预测情绪分析和众包映射技术来增强SA;（2）应用边缘计算技术和威胁监控工具来帮
analysis of social networks, and crowdsourced mapping techniques to enhance SA; (2) 助揭示潜在的行为模式，这些行为模式可能对人类分析师来说是"不可见的";（3）使用知识
表示，推理和ML算法来探索大型多维"决策空间"，以帮助开发知情和快速的"动态"替代行动
applying edge-computing technologies and threat monitoring tools to help reveal latent 方案。
patterns of behavior that are possibly “invisible” to human analysts alone; and (3) using
knowledge representation, reasoning, and ML algorithms to explore large multidimensional
“decision spaces” to help develop informed and rapid “on the fly” alternative courses of action.
Recognition of protected symbols. This is probably the most straightforward of the 受保护符号的识别。这可能是这里提出的最直接的应用程序，AI/ML方法用于识别指定受保
applications posited here, with AI/ML methods being used to identify accepted symbols for 护对象的接受符号，并相应地向操作员或命令链发出警报。受保护符号的存在并不意味着
designating protected objects and alerting the operator or the chain of command accordingly. 该地点实际上受到保护而不受攻击：该地点可能已经失去保护，或者不择手段的一方可能违
反国际法使用该符号来阻止攻击。但这种能力将提供一个安全网的情况下，受保护的符号存
The presence of protected symbols does not mean that the location is, in fact, protected from 在，但被操作部队错过。 145
attack: the location may have lost its protection or an unscrupulous party may be using the
symbol to deter attacks, in violation of international law. But this capability would provide a
safety net in cases when the protected symbol is present but was missed by operating forces.145
Alert on presence of transient civilians. One of the most frequent mechanisms for civilian 注意短暂平民的存在。在最近的美国行动中，平民伤害最常见的机制之一是在交战时将平
harm in recent US operations was the movement of civilians into the target area right around 民转移到目标地区。类似于Project Maven中使用的AI/ML对象识别功能可用于自动监视目
标区域周围的其他个体，并在检测到它们时发送警报。这一进程将提请行动部队注意这些
the time of the engagement. An AI/ML object-identification functionality akin to that used in
人,否则这些行动部队可能会盯着目标而错过短暂的平民存在。
Project Maven could be used to automatically monitor for additional individuals around the
target area and send an alert when they are detected. This process would bring such
individuals to the attention of operating forces that may otherwise fixate on the target and miss
Alert change from collateral damage estimate. AI/ML methods could be used to find 从附带损害估计警报变化. AI/ML方法可用于查找用于确定附带损害估计的图像与支持参与
differences between imagery used to determine the collateral damage estimate and more 的最近图像之间的差异。这个过程可以帮助确定行动部队可能无法识别的小细节，但这可
recent imagery taken in support of an engagement. This process can help identify little details 能是意外的平民存在的线索，例如建筑物附近的额外车辆。
that operating forces might not recognize but that could be cues of unanticipated civilian
presence, such as additional vehicles near a building.
Misassociation alert. Steps can also be taken to help identify that a misassociation has taken 误关联警报。还可以采取步骤来帮助确定发生了误关联。例如，应用程序可以识别被跟踪
place. For example, applications could recognize that a vehicle being tracked is not the same 的车辆不同
145We note that the Australian Armed Forces have recognized this application as a promising one and have 145我们注意到澳大利亚武装部队已经认识到这一应用是有前途的，并且已经进行了实地试验，显示了
already conducted field experiments showing the utility of this function. 这一功能的效用。
50 50
one that was being tracked previously, showing that a swap has occurred between a threat 一个以前被追踪的，显示威胁车辆和民用车辆之间发生了交换。
vehicle and a civilian vehicle.
A drone strike in Kabul: Could AI have helped? 喀布尔的无人机袭击：AI可以提供帮助吗？

On August 29, 2021, the US announced that it had used a drone to strike the vehicle of an ISIS- 2021年8月29日，美国宣布在阿富汗喀布尔使用无人机袭击了一名ISISK自杀式炸弹袭击者
K suicide bomber in Kabul, Afghanistan, averting an imminent threat to the airport during an 的车辆，避免了在空中撤离该国期间对机场的迫在眉睫的威胁。然而，很快，新闻报道表
airborne evacuation of the country. Soon, however, news reports suggested that the bomber 明，轰炸机不是唯一的受害者：一个家庭也出现在院子里，在爆炸中不幸遇难。随后的报
告显示，还有另一个根本错误：车辆驾驶员实际上不是自杀式炸弹袭击者，而是在美国国际
was not the only casualty: a family was also present in the courtyard and was tragically killed 开发署注册的人道主义组织的援助人员。 146
in the blast. Subsequent reporting showed that there was another fundamental error: the
driver of the vehicle was not in fact a suicide bomber but rather an aid worker with a
humanitarian organization registered with the US Agency for International Development.146
Although the event is being investigated by the US military, the causes of the incident (from 尽管美国军方正在调查这一事件，但事件的原因（来自我们在图7中的原因框架）似乎包括
our framework of causes in Figure 7) appear to include both misidentification (based on 错误识别（基于外观和行为）和短暂平民存在的附带损害。在这场悲剧发生后，值得问一
appearance and behavior) and collateral damage from the presence of transient civilians. In 问：ai支持的应用程序是否有所帮助？如果是这样，如何？
the aftermath of this tragedy, it is worthwhile to ask: could AI-enabled applications have
helped? And if so, how?
We can see several possible mitigation steps from Figure 7. For example, alerting the presence 我们可以从图7中看到几个可能的缓解步骤。例如，对临时平民的存在发出警报可能会给
of transient civilians may have given decision-makers additional time to scrutinize the 决策者更多的时间来仔细检查交战区域。这个警报可以与功能相结合，以实现战术耐心，
engagement area. This alert could have been combined with functions to enable tactical 也许是根据检测到的可能包括短暂平民在内的因素自动暂停参与。另一个AI应用程序可以
是开发更强大的民用生活模式的功能，为将其描述为威胁的车辆所做的假设提供额外的背景
patience, perhaps an automated pause to an engagement in light of detected factors that could 。例如，据报道，这辆车在遭到袭击前一天驶向人道主义组织所在地，作为其停靠点的一
include transient civilians. Another AI application could be a functionality to develop a more 部分。尽管美国军方认为访问可疑的威胁地点是其目标发展和生活模式的一部分，但没有
robust civilian pattern of life to provide additional context to the assumptions made about the 伴随的平民生活模式过程来试图确定可能显示个人或车辆不构成威胁的地点。
vehicle that described it as a threat. For example, reportedly the vehicle drove to the
humanitarian organization’s location as part of its stops during the day before it was attacked.
Although the US military considers visits to suspected threat locations as part of its target
development and pattern of life, there is no accompanying civilian pattern of life process to try
to identify locations that could show an individual or vehicle is not a threat.
146 Phil Stewart and Idrees Ali, “U.S. Says Kabul Drone Strike Killed 10 civilians, Including Children, in ‘Tragic 146Phil Stewart和Idrees Ali，"美国说喀布尔无人机袭击在'悲惨的错误'中杀死了包括儿童在内的10平民，"
Mistake,’” Reuters, Sept. 18, 2021, https://www.reuters.com/world/asia-pacific/us-military-says-10-civilians- 路透社，9月9日。 2021年18月https://www.reuters.com/world/asia-pacific/us-military-says-10-civilians
killed-kabul-drone-strike-last-month-2021-09-17/. killed-kabul-drone-strike-last-month-2021-09-17/.
51 51
Conclusions Conclusions
Although militaries speak of capabilities that help mitigate civilian harm, such as precision- 虽然军方谈到有助于减轻平民伤害的能力，如预埋弹药，但这些能力是为了更有效地打击
guided munitions, those capabilities were acquired to engage military targets more effectively. 军事目标而获得的。在某些情况下，它们在减轻伤害方面可能会有一些好处，但不是全部
They may have some benefit in mitigating harm in some circumstances but not all. For example, 。例如，当平民被误认为军事目标，而精确制导的弹药却误认为是这种错误的信念时，就
没有减轻平民伤害的价值。我们没有看到世界各地的军队寻求基于其在减轻平民伤害方面
a precision-guided munition has no value in mitigating civilian harm when civilians have been 的价值的野战能力。
misidentified as a military target and the munition is engaged in that mistaken belief. We do
not see militaries around the world seeking to field capabilities based on their value in
mitigating civilian harm.
The emerging technology of AI presents an opportunity for militaries to pursue this goal: AI- 人工智能的新兴技术为军队实现这一目标提供了一个机会：人工智能和其他降低平民战争
enabled and other applications for reducing the cost of war on civilians are within the realm of 成本的应用都在可能的范围内。无论是考虑到法律承诺，还是出于对尽一切可能使平民在
the possible. Both in light of legal commitments and out of interest in doing everything possible 发动战争时免受伤害的兴趣，各国都应该问自己：我们如何利用人工智能来保护平民免受伤
害？我们如何才能减少武装冲突期间造成的痛苦、伤害和破坏？
to spare civilians from harm in the waging of war, states should be asking themselves: How can
we use AI to protect civilians from harm? And how can we lessen the infliction of suffering,
injury, and destruction overall during armed conflict?
In this report we have provided an initial framework and methodology for identifying AI 在这份报告中，我们提供了一个初步的框架和方法，以确定人工智能应用，以帮助减轻平
applications to help mitigate civilian harm. For example, based on our analysis of particularly 民伤害。例如，基于我们对减少可用于人工智能应用的平民伤害的特别有益的缓解措施的
beneficial mitigation steps for reducing harm to civilians that are amenable to AI applications, 分析，我们建议以下功能作为有希望的起点:
we suggest the following functions as promising starting points:
• Alerting the presence of transient civilians by using object identification to • 通过使用物体识别自动监测目标区域周围的其他个人，并在检测到他们时发出警报

automatically monitor for additional individuals around the target area and send an ，提醒瞬态平民的存在。这项申请将提请行动部队注意这些个人,否则这些部队可
alert if they are detected. This application would bring these individuals to the 能会盯着目标而错过短暂的平民存在。
attention of operating forces that may otherwise fixate on the target and miss
• Detecting a change from collateral damage estimate by finding differences • 通过发现用于确定附带损害估计的图像与支持参与的最近图像之间的差异来检测附
between imagery used to determine the collateral damage estimate and more recent 带损害估计的变化。此应用程序可以帮助识别操作部队可能无法识别的小细节，但
imagery taken in support of an engagement. This application can help identify little 可能是意外的平民存在的线索，例如建筑物附近的额外车辆。
details that operating forces might not recognize but that could be cues of
unanticipated civilian presence, such as additional vehicles near a building.
• Alerting a potential miscorrelation by helping to identify that a miscorrelation has • 通过帮助确定发生了误关联，提醒潜在的误关联。例如，应用程序可以识别被跟
taken place. For example, applications could recognize that a vehicle being tracked is 踪的车辆与以前被跟踪的车辆不同，这表明威胁车辆和民用车辆之间发生了交换。
not the same one that was being tracked previously, showing that a swap has occurred
between a threat vehicle and a civilian vehicle.
• Recognizing protected symbols by using AI/ML methods to identify accepted • 通过使用AI/ML方法识别指定受保护对象（例如红十字或红新月）和警报的接受符
symbols for designating protected objects (e.g., red cross or red crescent) and alerting 号来识别受保护符号
52 52
the operator or the chain of command accordingly. The presence of protected symbols 操作员或相应的指挥链. 受保护符号的存在并不意味着该地点实际上受到保护而不
does not mean that the location is, in fact, protected from attack: the location may have 受攻击：该地点可能已经失去保护，或者不择手段的一方可能违反国际法使用该符
lost its protection or an unscrupulous party may be using the symbol to deter attacks, 号来阻止攻击。但这种能力将提供一个安全网的情况下，受保护的符号存在，但被
操作部队错过。 147
in violation of international law. But this capability would provide a safety net in cases
where the protected symbol is present but was missed by operating forces.147
Finally, by examining a recent tragic civilian harm incident in Afghanistan, we found we could 最后，通过审查阿富汗最近发生的悲惨平民伤害事件，我们发现我们可以从我们的根本原
identify mechanisms that led to civilian harm from our root cause framework and thus were 因框架中确定导致平民伤害的机制，从而能够从我们的AI应用矩阵中看到可能有助于避免该
able to see potential AI solutions from our AI applications matrix that could have helped to 事件中平民伤害的潜在AI解决方案。这种应用对这一现实世界事件的潜在效用表明，人工
智能应用可以以具体和实际的方式应用，以帮助减少（如果不是消除）武装冲突中的平民伤
avert civilian harm in that incident. The potential utility of such applications for this real world 害。
incident illustrates the possibility that AI applications can be applied in specific and practical
ways to help reduce, if not eliminate, civilian harm in armed conflict.
147We note that the Australian Armed Forces have recognized this application as a promising one and have 147我们注意到澳大利亚武装部队已经认识到这一应用是有前途的，并且已经进行了实地试验，显示了
already conducted field experiments showing the utility of this function. 这一功能的效用。
53 53
Appendix: AI/ML Approaches, Appendix: AI/ML Approaches,
Methods, and Algorithms Taxonomy 方法和算法分类
In this report (e.g., Figure 7) we discuss 33 basic classes of AI and ML applications organized 在本报告（如图7）中，我们讨论了33个基本类的AI和ML应用程序，这些应用程序被组织
into 11 “top-level” domains. Although the lists contain most of the key technologies that may 成11个"顶级"域。尽管这些清单载有可用于保护平民和减少平民伤害的大多数关键技术，
be used to protect civilians and reduce civilian harm and are based partly on how AI and ML 并且部分基于AI和ML在NSCAI最终报告中的特征，148本报告中提供的清单既不确定也不完
整。理想情况下，图3和图7将伴随第三个矩阵，将应用程序与特定方法相关联。为了权宜
are characterized in the NSCAI’s final report,148 the list provided in this report is neither 之计（因为我们离本文的重点太远，无法为建立这种明确的关联提供适当的背景），我们将
definitive nor complete. Ideally, Figure 3 and Figure 7 would be accompanied by a third matrix 讨论限制在总结AI/ML方法分类法的雏形上。我们借鉴了之前在2020年对最先进方法的调
that associates applications with specific methods. For expediency’s sake (because it would 查中发表的材料。 149
take us too far afield from the focus of this paper to provide the proper context for forging such
an explicit association), we limit our discussion to summarizing the rudiments of what an
AI/ML methodological taxonomy looks like. We draw from material previously published in a
2020 survey of state-of-the-art methods.149
Figure 8. The Five Tribes of AI Functions 图8。 AI功能的五个部落
148 NSCAI, Final Report. 148NSCAI,最后报告.

149 Ilachinski, Artificial Intelligence: Emerging Themes, Issues, and Narratives. 149Ilachinski,Artificial Intelligence:Emerging Themes,Issues,And Narratives.
54 54
Figure 8 is a schematic illustration of the "Five Tribes" of AI:150 (1) Bayesian approaches, which 图8是人工智能"五个部落"的示意图：150（1）贝叶斯方法，它依赖于通过似然估计的概率
rely on probabilistic inferences via likelihood estimates; (2) symbolist approaches, which are 推论；（2）象征主义方法，它可以追溯到20世纪80年代和90年代的"良好的老式人工智能"
throwbacks to the 1980s’ and 1990s’ “Good Old-Fashioned AI” and rely on logic and symbolic ，并依赖于逻辑和符号规则；（3）类比方法，主要由经典的函数优化和非线性分类方法组
成；（4）进化方法，它受到生物进化的启发（其中问题的"解决方案"甚至智能本身就是"）
rules; (3) analogical approaches, which consist largely of classical function optimization and 。生长"或使用基本的自然进化过程（如重组，交叉和突变）进化;和（5）连接主义方法，
nonlinear classification methods; (4) evolutionist approaches, which are inspired by biological 这是可以说是当今最流行的一类深度神经网络学习技术的另一个标签。从人工智能最顶层
evolution (wherein “solutions” to a problem⎯even intelligence itself⎯are “grown” or evolved 的"五个部落"分解中向下钻一个层次，就会发现一系列特定的方法、函数和算法（图8右侧
using basic natural evolutionary processes such as recombination, crossover, and mutation); 的列表）（故意地）太小--太难理解）。这个按字母顺序排列的列表的要点是，尽管它看
and (5) connectionist approaches, which is another label for what is arguably today’s most 起来很长，但它只包含现存AI方法的一小部分！
popular class of deep neural network learning techniques. Drilling down even a single level
from AI’s topmost “Five Tribes” decomposition reveals a (deliberately) too-small-to-
comfortably-read litany of specific methods, functions, and algorithms (the list that appears on
the right-hand side of Figure 8). The takeaway point of this alphabetized list is that, despite its
apparent length, it contains only a small subset of extant AI methods!
AI, having started in the 1950s, includes (but is more general than) the ML focus that ensued AI从20世纪50年代开始，包括（但比更普遍）20世纪80年代随之而来的ML重点，这反过
in the 1980s, which in turn is more general than the DL techniques currently in fashion. 来又比目前流行的DL技术更普遍。 "机器学习"是一个流行的短语，指的是各种技术，旨在
“Machine learning” is a catchall phrase that refers to a wide variety of techniques designed to 检测模式，学习和预测数据。具体技术包括如下：151
detect patterns in and learn and make predictions from data. Specific techniques include the
following:151
• Bayesian belief networks, which are graph models whose nodes represent some • 贝叶斯信念网络，这是图模型，其节点表示系统的某些对象或状态，其链接表示这
objects or states of a system and whose links denote probabilistic relationships among 些节点之间的概率关系
those nodes
• Deep learning, which refers to a class of ML algorithms designed to find multiple high • 深度学习，指的是一类ML算法，旨在查找数据中模式的多个高级抽象表示
levels of abstract representations of patterns in data
• Genetic algorithms and other evolutionary programming techniques that mimic the • 遗传算法和其他模仿自然选择动力学的进化编程技术152
dynamics of natural selection152
• Inductive logic programming, designed to infer a hypothesis from a knowledge base • 归纳逻辑编程，旨在从知识库和一组正面和负面的例子中推断出一个假设153
and a set of positive and negative examples153
150The “Five Tribes” decomposition is borrowed from Pedro Domingos, The Master Algorithm: How the Quest for 150"五个部落"的分解是从Pedro Domingos，The Master Algorithm：How The Quest for The Ultimate Learni
the Ultimate Learning Machine Will Remake Our World, (New York: Basic Books, 2015). ng Machine Will Remake Our World，（New York：Basic Books，2015）借来的。
151 Russell and Norvig, Artificial Intelligence: A Modern Approach. 151Russell and Norvig,Artificial Intelligence:A Modern Approach.
Zbigniew Michalewicz and David B. Fogel, How to Solve It: Modern Heuristics, (New York: Springer-Verlag,
152 152Zbigniew Michalewicz和David B.Fogel，如何解决它：现代启发式，（纽约：Springer-Verlag，2005
2005). ）。
Stephen H. Muggleton and Hiroaki Watanabe, eds., Latest Advances in Inductive Logic Programming, (London:
153 153Stephen H.Muggleton和Hiroaki Watanabe，eds。，Latest Advances in Inductive Logic Programming，
Imperial College Press, 2014). （伦敦：帝国理工学院出版社，2014）。
55 55
• Neural networks, which are inspired by the structure and function of biological neural • 神经网络，其灵感来自生物神经网络的结构和功能154
networks154
• Reinforcement learning, which is inspired by behavioral psychology and refers to a • 强化学习，它受到行为心理学的启发，指的是一种技术，通过适应性地构建一系列
technique whereby learning proceeds by adaptively constructing a sequence of 行动来学习，这些行动共同最大化一些长期奖励155
actions that collectively maximize some long-term reward155
• Support vector machines, which are essentially multidimensional binary classification • 支持向量机，其本质上是多维二进制分类算法156
algorithms156
Although all ML techniques require a dataset (or multiple datasets) as a source of training data, 尽管所有ML技术都需要一个数据集（或多个数据集）作为训练数据的来源，但学习可以通
the learning can proceed in one of three ways: supervised, semi-supervised, or unsupervised. 过以下三种方式之一进行：监督、半监督或无监督。在监督学习中，每个训练数据元素都
被明确标记为输入-输出对，其中输出是希望系统学习与给定输入相关联的"正确"期望值（
In supervised learning, each training data element is explicitly labeled as an input-output pair,
从而学习将原始训练集中没有的输入-输出对相关联的一般规则），"输出"表示"监督信号"
where the output is the “correct” desired value that one wishes the system to learn to associate 。"在无监督学习中，系统试图自行发现数据中的隐藏结构-也就是说，在系统处理训练数据
with a given input (thereby learning the general rules by which to associate input-output pairs 时，不会给出奖励信号来"推动"系统。半监督学习是指也使用未标记训练数据的一类监督
not in the original training set) and the “output” represents a “supervisory signal.” In 学习技术。强化学习可以被认为是半监督学习的一种形式，因为它既不使用输入-输出对进
unsupervised learning, the system attempts to discover hidden structure in data on its own— 行训练，也不完全无监督;相反，系统接收的反馈类型取决于其响应。对于正确的响应，它
that is, no reward signals are given to “nudge” the system as it processes the training data. 会收到与任何监督学习系统相同类型的响应（例如，响应是"正确的"）;对于不正确的响应
，它被告知给出了"不正确的响应"，但没有被告知正
Semi-supervised learning refers to a class of supervised learning techniques that also use
unlabeled training data. Reinforcement learning may be considered a form of semi-supervised
learning in that it neither uses input-output pairs for training nor is completely unsupervised;
instead, the type of feedback the system receives depends on its response. For correct
responses, it receives the same type of response as any supervised learning system does (e.g.,
response is “correct”); for incorrect responses, it is told that an “incorrect response” was given
but is not informed of what the correct response was.
Figure 9 shows a top-level view of a taxonomy of ML organized by functional similarity, 图9示出了由功能相似性组织的ML的分类学的顶层视图，其中算法的每个主要簇伴随有视
wherein each major cluster of algorithms is accompanied by a visual schematic that serves as 觉示意图，其作为给定类别算法被设计用来做什么的助记符参
a mnemonic reference for what a given category of algorithms is designed to do.
The mind map in Figure 10 contains the full high-resolution version of this taxonomy and 图10中的思维导图包含此分类法的完整高分辨率版本，包括100多个特定算法以及对主要
includes 100-plus specific algorithms and embedded hot-link references to primary and 和次要参考源的嵌入式热链接引用。
secondary reference sources.
154 Mohamad Hassoun, Fundamentals of Artificial Neural Networks, (Cambridge, MA: MIT Press, 2003). 154Mohamad Hassoun,Fundamentals Of Artificial Neural Networks,(Cambridge,MA:MIT Press,2003)。
155Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction, (Cambridge, MA: MIT Press, 155Richard S.Sutton和Andrew G.Barto,Reinforcement Learning:An Introduction,(Cambridge,MA:MIT Press,1
1998). 998)。
156Nello Cristianini, An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods, (New 156Nello Cristianini,An Introduction to Support Vector Machines and Other Kernel-Based Learning Methods
York: Cambridge University Press, 2000). ,(New York:Cambridge University Press,2000)。
56 56
Figure 9. High level taxonomy of machine learning methods 图9。机器学习方法的高级分类法
57 57
Figure 10. Full taxonomy of machine learning methods 图10。机器学习方法的全分类法
See next page. 请参阅下一页。
58 58
General class of algorithms inspired by the structure and function of the brain 受大脑结构和功能启发的一般算法类
Equivalent to linear regression (i.e., linearly seperable problems) Perceptron 等价于线性回归（即线性可分离问题） Perceptron
An associative memory network Fully connected neural network with a symmetric matrix of links Hopfield Network (HN) 关联记忆网络具有对称链接矩阵的全连接神经网络 Hopfield Network (HN)
Typically 3-layer network that uses radial basis functions as activation functions Radial Basis Function Network (RBFN) Self-defining: defines the class of problem and class of methods 通常使用径向基函数作为激活函数的3层网络径向基函数网络(RBFN) 自定义：定义问题的类和方法的类
Though it hardly does justice to the enormous space of DL architectures, generally speaking, a DL network is an ANN with more than two layers k-Means: simplest unsupervised learning algorithm Calculates the centroids of k clusters and assigns a data point to that cluster having least distance between its centroid and the data point 尽管DL架构的巨大空间并不合理，但一般来说，DL网络是一个包含两层以上的ANN k-Means：最简单的无监督学习算法计算k个聚类的质心，并将一个数据点分配给该聚类，该聚类的质心与数据点之间的距离最小
Numerous other algorithms AlexNet, GoogLeNet, LeNet, ResNet, VGGNet, ZFNet k-Medians Variation of k-Means: rather than calculating the mean for each cluster (to define its centroid), calculates the median 许多其他算法 AlexNet, GoogLeNet, LeNet, ResNet, VGGNet, ZFNet 卷积神经 k-Medians K均值的变化：不是计算每个聚类的均值（以定义其质心），而是计算中位数
Convolutional Neural
Optimized for visual processing tasks Architecture inspired by organization of visual cortex ("receptive fields") Network (CNN) Expectation Maximisation (EM) Searches across probability distributions and their parameters 针对视觉处理任务进行了优化受视觉皮层组织启发的建筑（"接受领域"）
Network (CNN) Expectation Maximisation (EM) 跨概率分布及其参数的搜索
Decision flow is unidirectional, advancing from the input to the output in successive layers, without cycles or loops Multilayer Perceptrons (MLP) / Feed-Forward Neural Network (FFNN) Fuzzy C-means (FCM) Data point assigned a probability of belonging to cluster 决策流是单向的，在连续的层中从输入推进到输出，没有循环或循环多层感知器(MLP)/前馈神经网络(FFNN) Fuzzy C-means (FCM) 数据点分配了属于集群的概率
Specifically, remember things learned from prior input(s) while generating output(s) Designed so that decisions are influenced by what is "remembered" from past via loops Recurrent Neural Networks (RNNs) Agglomerative: bottom-up approach Initially each data point is considered as an individual cluster Iteratively, similar clusters merge with other clusters until one cluster or K clusters are formed Numerous implementations 具体来说，在生成输出时记住从以前的输入中学到的东西设计使决策受到通过循环从过去"记住"的内容的影响循环神经网络(Rnn) Agglomerative: bottom-up approach 最初，每个数据点都被视为单个集群迭代地，类似的聚类与其他聚类合并，直到形成一个聚类或K个聚类 Numerous implementations
Hierarchical Clustering (HC) 层次聚类(HC)
Well-suited to classifying, processing and making predictions based on time series data Capable of learning long-term dependencies A special kind of RNN Long Short-Term Memory Networks (LSTMs) Clustering Divisive: top-down approach All data points are treated as a single cluster in each iteration Then separate the data points from a cluster that are not similar Numerous implementations 非常适合基于时间序列数据进行分类、处理和预测能够学习长期依赖关系一种特殊的RNN 长期短期记忆网络(LSTMs) Clustering Divisive: top-down approach 在每次迭代中，所有数据点都被视为单个群集然后将数据点与不相似的群集分开 Numerous implementations
Denoising Numerous other algorithms See Deep Learning... See Dimensionality Reduction... Denoising 许多其他算法请参阅深度学习。.. 请参见降维。..
Sparse Learns to compress and encode data (in latent space) and how to Sparse 学习压缩和编码数据（在潜在空间中），以及如何将数据重建回尽可能接
Autoencoder 近原始输入 Autoencoder
reconstruct the data back to being as close to the original input as possible
Stacked Auto-Encoders Seeks to leverage any inherent structure in a dataset in order to summarize or describe data using less information Stacked Auto-Encoders 寻求利用数据集中的任何固有结构，以便使用较少的信息汇总或描述数据
Variational Autoencoders (VAE) Autoencoder Learns to compress and encode data (in latent space) and how to reconstruct the data back to being as close to the original input as possible Variational Autoencoders (VAE) Autoencoder 学习压缩和编码数据（在潜在空间中），以及如何将数据重建回尽可能接近原始输入
An unsupervised, probabilistic, model with a complete set of undirected connections between different layers Deep Boltzmann Machine (DBM) 1st PC captures direction of maximum variability in the data 一个无监督的概率模型，具有不同层之间的一整套无向连接深玻尔兹曼机(DBM) 第一台PC捕获数据中最大变异性的方向
Nodes in each layer are connected to all the nodes in the previous and subsequent layer Designed so that each layer learns the entire input Deep Belief Networks (DBN) / Bayesian Belief Network (BBN) Principal Component Analysis (PCA) 2nd PC captures remaining variance but has variables uncorrelated with 1st PC
Maps maximum variance in data to new coordinate system with axes = ‘principal components’ (PCs) 每个层中的节点都连接到上一层和下一层中的所有节点设计为使每个层学习整个输入深度信念网络(DBN)/贝叶斯信念网络(BBN) 主成分分析(PCA) 将数据的最大方差映射到具有轴='主分量'(PCs)的新坐标系第二台PC捕获剩余方差，但变量与第一台PC不相关
Balances interpretability and accuracy Modified hierarchical classifiers that use trees constructed in weight-space Neural-Backed Decision Trees (NBDT) All successive PCs capture remaining variance but are uncorrelated with previous PC 平衡可解释性和准确性修改的分层分类器，使用在权重空间中构造的树神经支持决策树(NBDT) 所有连续的Pc都捕获剩余的方差，但与以前的PC不相关
Deep Learning (DL) 深入学习(DL)
Uses discrete activation thresholds ("spikes") rather (Attempt to) bridge the gap between neuroscience and ML by Data variables assumed to be linear mixtures of unknown latent variables 使用离散激活阈值（"尖峰"），而不是像"常规"Nn （试图）通过使用类似于我们大脑神经元功能的模型来弥合神假定为未知潜在变量的线性混合的数据变量
Spiking Neural Networks (SNN) 螺旋神经网络(SNN) 独立组件分析观测多元数据的生成模型
than continuous values as in "conventional" NNs using models similar to how the neurons in our brain function Independent Component Analysis A generative model for observed multivariate data 那样的连续值经科学和ML之间的差距
The latent variables (i.e., "independent components") are assumed non-gaussian and mutually independen 潜在变量（即"独立分量"）被假定为非高斯和相互独立
Allows for much more parallelization than RNNs and therefore reduced training time Does not require that sequential data be processed in order 允许比Rnn更多的并行化，因此减少了训练时间不要求按顺序处理顺序数据
Principal Component Regression (PCR) Designed to reduce multi-collinearity of a dataset PCR (effectively) uses PCA, then linear regression on new "principal components" (PCs) 主成分回归(PCR) 旨在减少数据集的多共线性 PCR（有效）使用PCA，然后对新的"主成分"（PCs）进行线性回归
SOTA for many NLP problems Leverages "attention" without using recurrent sequential processing (as by RNNs and LSTMs) Artificial Neural 许多NLP问题的SOTA 利用"注意力"而不使用循环顺序处理（如Rnn和LSTMs）人工神经网络(AN
Partial Least Squares Regression (PLSR) Finds a linear regression model by projecting predicted variables and the observable variables to a new space 偏最小二乘回归(PLSR) 通过将预测变量和可观察变量投影到新空间来查找线性回归模型
1st deeply bidirectional, unsupervised language representation, pre-trained using only a plain text corpus (i.e., Wikipedia). Bidirectional Encoder Representations from Transformers (BERT) Networks (ANN) 第一个深度双向，无监督的语言表示，只使用纯文本语料库（即维基百科）进行预训练。来自变压器的双向编码器表示(BERT) N)
Transformer 变压器架构
Sammon Mapping Basic technique (from 1969): a nonlinear mapping from a high-dimensional space to a lower-dimensional space while preserving the inter-point distances in high-dimensional space in the lower-dimension projection Sammon Mapping 基本技术（从1969年开始）：从高维空间到低维空间的非线性映射，同时在低维投影中保留高维空间中的点间距离
Architecture
GPT-3 GPT-2 Numerous implementations GPT-3 GPT-2 Numerous implementations
Generative Pre-trained Transformer (GPT) Multidimensional Scaling (MDS) 生成预训练变压器(GPT) 多维缩放(MDS)
(175B Parameters) (1.5B Parameters) A nonlinear mapping of "information about pairwise 'distances' among N objects" into a configuration of N points in an abstract Cartesian space (175B Parameters) (1.5B Parameters) 在抽象笛卡尔空间中，"关于n个物体之间两两'距离'的信息"到N个点的配置的非线性映射
Dimensionality Reduction 降维查找表征或分隔两个或多个对象或事件类的要素的线假设预测变量是正态分布的，并且类具有相同的方差（或多变量问题的相同协
Finds linear combinations of features that characterize Assumes that predictor variables are normally distributed and the classes have 线性判别分析(lda)
Linear Discriminant Analysis (LDA) 性组合方差矩阵）
GNNs have similar properties to DL counterparts: capable of regression, classification, along with generating data for empty nodes, edges or separate two or more classes of objects or events identical variances (or identical covariance matrices, for multivariate problems) Gnn具有与DL对应物相似的属性：能够回归，分类，以及为空节点，边生成数据
Aim to learn node representations with recurrent neural architectures Pioneer GNN methods Recurrent graph neural networks (RecGNNs) Quadratic Discriminant Analysis (QDA) Does not assume (as does LDA) that the covariance matrix of classes is the same 目的学习循环神经架构的节点表示先驱GNN方法循环图神经网络(Recurrent graph neural networks,RecGNNs) 二次判别分析(QDA) 不假设（LDA也是如此）类的协方差矩阵是相同的
Generalize the operation of convolution from grid data to graph data

Generate a node V’s representation by aggregating its own features f(V) and neighbours’ features f(U) Convolutional graph neural Mixture Discriminant Analysis (MDA) Essentially, LDA but where it is assumed that each class is a Gaussian mixture of subclasses 通过聚合自己的特征f(V)和邻居的特征f(U)生成节点V的表示格数据到图数据的卷积的操作泛化卷积图神经网络(ConvGNN 混合判别分析(MDA) 本质上，LDA但其中假设每个类是子类的高斯混合
Unlike RecGNNs, ConvGNNs stack multiple graph convolutional layers to extract high-level node representations networks (ConvGNNs) s) 灵活判别分析(FDA) 使用预测变量的非线性组合（例如，样条）
Flexible Discriminant Analysis (FDA) Non-linear combinations of predictors is used (e.g., splines) 与RecGNNs不同，ConvGNNs堆叠多个图卷积层以提取高级节点表示
Unsupervised learning frameworks which encode nodes/graphs into a latent Generalized Discriminant Analysis (GDA) Maps a dataset from a high-dimensional space into a low-dimensional space by maximizing the ratio of between-class scatter to within-class scatter 无监督学习框架，将节点/图编码到潜在向量空间，并从编码信息中重建图数广义判别分析(gda) 通过最大化类间散点与类内散点的比率，将数据集从高维空间映射到低维空间
Learn network embeddings and graph generative distributions Variational Graph Autoencoders (VGAEs) 学习网络嵌入和图形生成分布 Variational Graph Autoencoders (VGAEs)
vector space and reconstruct graph data from the encoded information 据
Models each high-dimensional object by a 2 or 3 dimensional point in such a way that similar objects 通过2或3维点对每个高维物体进行建模，这样类似的物体由附近的点建模，不同的物体由远处的点建模
Consider spatial dependency and temporal dependency at the same time Aim to learn hidden patterns from spatial-temporal graphs Spatial-temporal graph neural networks (STGNNs) T-Distributed Stochastic Neighbor Embedding (t-SNE) 在同一时间考虑空间依赖性和时间依赖性目的从时空图中学习隐藏模式时空图神经网络(STGNNs) T分布随机邻域嵌入(t-SNE)
are modeled by nearby points and dissimilar objects are modeled by distant points with high probability ，概率很高
Numerous algorithms Adversarial Training 众多算法对抗性训练对图形应用对抗性
Apply adversarial 图对抗性方法
Graph Adversarial Methods 方法
Numerous algorithms Adversarial Attack methods to graphs Graph Neural ("Memory based") Compare new problem 'instances' with instances seen in training; similarity measures used to find best match and make prediction 众多算法对抗性攻击图神经（"基于内存"）将新问题"实例"与训练中看到的实例进行比较;用于查找最佳匹配并进行预测的相似性度量
Numerous algorithms Graph generation, prediction of nodes, node-pairs, link, labels Graph Reinforcement Learning (GRL) Networks (GNNs) 众多算法图生成、节点预测、节点对、链接、标签图强化学习(GRL) Networks (GNNs)
k-Nearest Neighbors (kNN) Parses entire dataset to find k-nearest instances to the new instance k-Nearest Neighbors (kNN) 解析整个数据集以查找与新实例最近的k个实例
Concept for how GNNs can support relational reasoning and combinatorial generalization, laying the foundation for more sophisticated, interpretable, and flexible patterns of reasoning Graph Relational Networks (Framework) Learning Vector Quantization (LVQ) (A kind of ANN) Unlike kNN, LVQ allows one to choose how many training instances to retain and learns exactly what those instances should look like Precursor to Self-Organizing Map (SOM)
GNNs如何支持关系推理和组合泛化的概念，为更复杂、可解释和灵活的推理模式奠定基础图关系网络（框架）学习矢量量化(LVQ) （一种ANN）与kNN不同，LVQ允许人们选择保留多少个训练实例，并确切地了解这些实例应该是什么样子自组织地图（SOM）的前身
GNNs have similar properties to DL counterparts: capable of regression, classification, along with generating data for empty nodes, edges Gnn具有与DL对应物相似的属性：能够回归，分类，以及为空节点，边生成数据
General class of methods designed to automate the process of applying machine learning Self-Organizing (A kind of ANN) Leverages competitive learning rather than error-correction (such as backpropagation with Since it typically produces a low-dimensional discretized representation 用于自动化应用机器学习算法的过程的一般类方法；通常涵盖从原始数据集到最终可部署模型 Self-Organizing （一种ANN）利用竞争学习而不是纠错（例如带有梯度下降的反向传播）：使用保留输入空间拓扑属性的邻由于它通常会产生输入空间的低维离散化表示，因此它作为"降维"方法
algorithms; generally covers the complete pipeline from raw dataset to final deployable model Map (SOM) gradient descent): use a neighborhood function that preserves the topological properties of the input space of input space, it is useful as a "dimensionality reduction" method 的完整管道 Map (SOM) 域函数很有用
Discovers complete ML algorithms using basic mathematical operations as building blocks AutoML-Zero 发现使用基本数学运算作为构建块的完整ML算法 AutoML-Zero
Instance-based ML产品套件，使具有有限机器学习专业知识的开发人员能够训练高质量模型
Instance-based
Suite of ML products that enables developers with limited machine learning expertise to train high-quality models Cloud AutoML Cloud AutoML
Effectively divides a global learning problem (such as tackled by 有效地划分全球学习问题（如
Numerous algorithms Data collection; cleaning; augmentation Data Preparation Locally Weighted Learning (LWL) / Memory-Based Learning (MBL) Typically by dividing cost function into multiple independent local cost functions 众多算法数据收集；清洁；增强数据准备局部加权学习(LWL)/基于内存的学习(MBL) 通常通过将成本函数划分为多个独立的局部成本函数
ANNs, SVM, and kNN)into multiple smaller/simpler learning problems ANNs，SVM和kNN）成多个更小/更简单的学习问题
Numerous algorithms Data collection; cleaning; augmentation Feature Engineering (FE) 众多算法数据收集；清洁；增强特征工程(FE)
Support Vector Machines (SVM) Binary classification algorithm Given a set of points of 2 types in N dimensional place, an SVM generates an (N - 1) dimensional hyperplane to separate those points into 2 groups 支持向量机(SVM) 二元分类算法给定一组在N维位置的2类型的点，SVM生成一个（N1）维超平面以将这些点分成2组
Numerous algorithms Hyperparameter optimization Hyperparameter Optimization (HPO) 众多算法 Hyperparameter optimization Hyperparameter Optimization (HPO)
Numerous algorithms Aims to improve the learning algorithm itself uusing multiple learning episodes Meta Learning 众多算法旨在利用多个学习集改进学习算法本身元学习指的是一般类的"stand-slone"计算模型，以及与更广泛的ML算法相关的核心类方法，其性能取决于最小化特定的损失/成本函数
Refers to both a general class of "stand-slone" computational models and as a core class of methods relevant to broader ML algorithms whose performance depends on minimizing a specific loss/cost function
Auto-WEKA; Hyperopt-Sklearn; Auto-sklearn; TPOT: A Tree-Based Pipeline Optimization Tool Auto-WEKA;Hyperopt-Sklearn;Auto-sklearn;TPOT:基于树的管道优化工具
Bayesian optimization Explicitly builds builds a probabilistic model 贝叶斯优化显式构建概率模型
Almost all algorithms are of a "trial, 几乎所有的算法都是"试验、修改、
modify, test, and update" variety Concerned with using various optimization algorithms to design optimal neural network architectures Classification-based optimization (CBO) Builds simple classifiers that divide the search space into positive and negative areas 测试和更新"的品种关注使用各种优化算法设计最优神经网络架构 Classification-based optimization (CBO) 构建简单的分类器，将搜索空间划分为正区域和负区域
Search space: dimensionality, complexity, etc. GA applied Similar to GP, but structure of the program to be optimized is Main variation 搜索空间：维度，复杂性等。 GA应用于程与GP类似，但待优化程序的结构是固定的，而其数值参数是主要变异算子是突
Evolutionary Programming (EP) 进化规划(EP)
Can categorize in to programs fixed, while its numerical parameters are allowed to evolve operator is mutation 可以进行三维分序允许演化的变
Strategy: how to explore; faces typical 'exploration-exploitation' trade-off 战略：如何探索；面临典型的"探索-开发"权衡
three dimensions Automated ML (AutoML) Generally uses asexual reproduction: new individuals typically produced without Concerned w/finding optimal ' self adaptive' adjustment
类 Automated ML (AutoML) 通常使用无性繁殖：新个体通常在没有交叉或类似技术的情况下产生，在遗传算法（关注w/寻找最佳的"自适应"调整突变大小以生产新样本
Performance estimation: standard training and evaluatioon often too costly crossover or similar techniques more prevalent in the field of genetic algorithms (GA) of the mutation sizes for the production of new samples 绩效评估：标准培训和评估通常成本过高 GA）领域更为普遍
Evolutionary Strategy (ES) （一种使用的开源方法）用于指导网络态射的贝叶斯优化
进化策略(ES) 协方差矩阵适应(cma)
(An open source method that uses) Bayesian optimization used to guide network morphism Auto-Keras Covariance Matrix Adaptation (CMA) Adaptively increases or decreases the search space for the next generation given the current generation’s fitness Auto-Keras 考虑到当前一代的适应度，自适应地增加或减少下一代的搜索空间
Evolutionary Evolutionary
Based on continuous relaxation of architecture representation, Numerous other algorithms 基于体系结构表示的连续松弛，允许使用梯度下降进行高效搜许多其他算法
Leverages leverages game theoretic concepts Differentiable Architecture Search (DARTS) Algorithms (EA) 利用博弈论概念
索
可微架构搜索(DARTS) Algorithms (EA)
allowing efficient search using gradient descent
Genetic Algorithms (GA) (Invented in 1960 by John Holland) Most general method using recombination, crossover and mutation) 遗传算法(GA) （由John Holland于1960年发明）使用重组，交叉和突变的最通用方法）
Meta-ML 具体来说，首先在CIFAR-10上搜索最佳卷积层或单元格，然后通过将此单元在小型数据集上搜索架构构建块，然后将该块传输到
Meta-ML
Specifically, search for best convolutional layer or cell on CIFAR-10 first, then Searches for an architectural building block on a small
NASNet Neural Architecture Search (NAS) Selection of fittest programs for reproduction (crossover) and mutation according NASNet 神经架构搜索(NAS) 根据预定义的适应度度量（通常，在期望的任务上的性能）选择用于再现（交叉
apply this cell to the ImageNet by stacking together more copies of this cell dataset and then transfer the block to a larger dataset Genetic Programming (GP) GA applied to programs 格的更多副本堆叠在一起，将此单元格应用于ImageNet 更大的数据集遗传编程(GP) GA应用于程序
to a predefined fitness measure (typically, performance on a desired task ）和突变的适者生存程序
Searches for structures of CNNs in order of increasing complexity Progressive Neural Architecture Search (PNAS) 以增加复杂性的顺序搜索CNNs的结构渐进神经架构搜索(PNAS)
Algorithms inspired by biological evolution 受生物进化启发的算法
See Reinforcement Learning... See Evolutionary Algorithms (EA).. Numerous other algorithms 请参阅强化学习。.. 请参阅进化算法(EA)。. 许多其他算法
Monte-Carlo Tree Search (MCTS) (in a game-playing context/ MCTS was introduced in 2006 for Go) Leverages analysis of the most promising moves, by expanding a search tree based on random sampling of the search space 蒙特卡罗树搜索(MCTS) （在游戏环境中/MCTS于2006年为Go引入）通过扩展基于搜索空间随机抽样的搜索树，利用对最有前途的移动的分析
Meta-algorithms that combine multiple ML methods into one predictive model Optimization 将多个ML方法组合成一个预测模型的元算法 Optimization
See Reinforcement learning... 请参阅强化学习。..
Fitting and inference are accomplished via an iterative backfitting Markov chain Monte Carlo (MCMC) algorithm Each tree is constrained by a prior to be a weak learner Bayesian Additive Regression Trees (BART) 拟合和推理是通过迭代反拟合马尔可夫链蒙特卡罗（MCMC）算法完成的每棵树都受到一个弱学习者之前的约束贝叶斯加性回归树(BART)
Simultaneous Optimistic optimization (SOO) A branch-and-bound approach A "blackbox" function optimization toolkit that finds global optimum of a function 同时乐观优化(SOO) A branch-and-bound approach 一个"黑盒子"函数优化工具包，可以找到函数的全局最优
Boosting done on "Decision Stump" (i.e., a "1 rule" ML model consisting of a one-level decision tree) Adaptive Boosting (AdaBoost) 在"决策树桩"（即由一级决策树组成的"1规则"ML模型）上完成自适应提升(AdaBoost)
Inspired by the behavior of biological communities that exhibit both individual and social behavior 受到表现出个人和社会行为的生物群落行为的启发
Open-source gradient boosting on decision trees CatBoost Particle Swarm Otimization (PSO) 决策树上的开源梯度提升 CatBoost 粒子群优化(PSO) 由一组候选解（"粒子"）组成，这些候选解（"粒子"）根据其位置和速度（以及与其他代理的
Consists of a population of candidate solutions ("particles") that move around in the search-
space according to functions of their position and velocity (and relationship with other agents) 关系）的函数在搜索空间中移动
LightGBM: uses tree based learning algorithms Gradient Boosting Machines (GBM) Boosting LightGBM：使用基于树的学习算法梯度提升机(GBM) Boosting
Swarm Intelligence (SI) Ant Colony Optimization (ACO)
Mimicks the way ants create and mark trails from
以弱预测模型的集合形式生成预测模型，通常是决策树梯度提升回归树(GBRT)
群体情报(SI) 蚁群优化(ACO)
模仿蚂蚁创造和标记从巢穴到食物来源的痕迹的
Yields prediction model in the form of an ensemble of weak prediction models, typically decision trees Gradient Boosted Regression Trees (GBRT) their nest to food sources and back again 方式，然后再回来
为速度和性能设计的梯度提升决策树的实现
Implementation of gradient boosted decision trees designed for speed and performance XGBoost
A Taxonomy of ML Methods and Functions Stochastic Diffusion Search (SDS) First SI metaheuristic introduced in 1989 XGBoost
ML方法和功能的分类随机扩散搜索(sds) 1989年首次引入SI元动脉学
Given a training datasetD of size N, bagging generates M new training sets, {Di}, each of sizeN', by sampling from D uniformly with replacement Bootstrapped Aggregation (Bagging) Methods that derive from the collective behavior of decentralized, self-organized systems (natural or artificial) 给定大小为N的训练数据集D，bagging生成M个新的训练集{Di}，每个大小为N'，通过从D均匀采样并替换引导聚合（装袋）源自分散，自组织系统（自然或人工）的集体行为的方法
Simplest method: each ensemble member contributes an equal amount to the final prediction Weighted Average (Blending) Ensembles 最简单的方法：每个集成成员对最终预测的贡献量相等加权平均(混合) Ensembles
Uses a new model to learn how to best combine the predictions from two or more models trained on your dataset Stacked Generalization (Stacking) A workhorse of basic statistics that has been co-opted by its modern incarnation as "statistical machine learning"; models the relationship between variables 使用新模型了解如何最好地组合在数据集上训练的两个或多个模型的预测堆叠泛化（Stacking）基础统计的主力，被其现代化身为"统计机器学习"所选择;对变量之间的关系进行建模
Step 1 - Bootstrap sampling: each generated training set composed of random subsamples from original dataset Least-Angle Regression (LARS) A SOTA algorithm to solve L1 regularized linear regression (LR) or logistic regression (LogR) 步骤1引导采样：每个生成的训练集由原始数据集的随机子采样组成 Least-Angle Regression (LARS) 求解l1正则化线性回归（lr）或逻辑回归（Logistic regression）的SOTA算法
(Bagging with) Random Forest （装袋）随机森林
Step 2 - Random forest: create multiple models by using the same algorithm on different generated training sets Linear Regression (LR) Arguably the simplest and best known method: assumes a linear relationship between observed and predicted variables Useful only for linearly separable problems 步骤2随机森林：在不同生成的训练集上使用相同的算法创建多个模型线性回归(lr) 可以说是最简单和最着名的方法：假设观察变量和预测变量之间存在线性关系仅对线性可分的问题有用
Locally Estimated Scatterplot Smoothing (LOESS) A non parametric best-fit curve of best fit to a dataset without assuming the data must fit a given distribution 局部估计散点图平滑（黄土）最佳拟合到数据集的非参数最佳拟合曲线，而不假设数据必须适合给定的分布
Multiple Regression (MR) Models a dependent variable as a function of several independent variables with corresponding coefficients, along with a constant term 用于学习规则的基于规则的方法，该方法最能描述数据集中变量之间的模式和关系多元回归(MR) 将因变量建模为几个具有相应系数的自变量的函数，以及一个常数项
Rule-based method for learning rules that best describe patterns and relationships among variables in a dataset
Typically used in a transactional database to mine frequent item sets and then generate association rules Apriori algorithm Multivariate Adaptive Regression Splines (MARS) A non-parametric technique that automatically models nonlinearities and interactions between variables 通常在事务数据库中用于挖掘频繁项集，然后生成关联规则 Apriori algorithm 多元自适应回归样条（MARS）一种非参数技术，可自动对变量之间的非线性和相互作用进行建模
Where Apriori proceeds in a horizontal sense (like the Breadth-First Search of a graph), ECLAT algorithm proceeds in a vertical manner (like Depth-First Search of a graph Equivalence Class Clustering and bottom-up Lattice Traversal (ECLAT) algorithm Regression Ordinary Least Squares Regression (OLSR) Simple linear method: "relationship" revealed by minimizing sum of squares of difference between observed and predicted variables 当Apriori以水平方式进行时（如图的广度优先搜索），ECLAT算法以垂直方式进行（如图的深度优先搜索等价类聚类和自下而上格遍历（ECLAT）算法 Regression 普通最小二乘回归（OLSR）简单线性方法：通过最小化观察变量和预测变量之间的差平方和来揭示"关系"
Represents a database in the form of a tree called a "frequent pattern" tree An improced variant of Apriori Frequent Pattern (FP) See Artificial Neural Networks (ANN)... 以称为"频繁模式"树的树的形式表示数据库 Apriori的改良变体频繁模式(FP) 请参阅人工神经网络(Ann)。..
Association Association
Numerous other algorithms 许多其他算法
Rules Rules
General class of methods of adding information in order to prevent overfitting 为了防止过度拟合而添加信息的一般类方法
Data augmentation (Easiest way to reduce overfitting is to) increase size of training data 数据增强（减少过拟合的最简单方法是）增加训练数据的大小
See Artificial Neural Networks ... 请参阅人工神经网络。..
Dropout Iteratively, randomly remove some nodes (really, an ensemble technique) Dropout 迭代地，随机删除一些节点（真的，一个整体技术）
A discriminative classifier yields the the conditional probability of the target Y, given observation x: P(Y |X=x) 判别分类器产生目标y的条件概率，给定观察x：P（Y/X=x）
Elastic Net Linearly combines the L1 and L2 penalties of the lasso and ridge methods 弹性网线性组合lasso和ridge方法的L1和L2惩罚
Explicitly construct a model of possible decisions made based on values of attributes in dataset 显式构建基于dataset中属性值做出的可能决策的模型
Performs both variable selection and regularization in order to enhance prediction accuracy and interpretability 执行变量选择和正则化，以提高预测精度和解释性
C4.5 and C5.0 (different versions of a powerful approach) Least Absolute Shrinkage and Selection Operator (LASSO) / L1 C4.5和C5.0（不同版本的强大方法）最小绝对收缩和选择运算符(LASSO)/L1
Regularization Similar to "Ridge Regression" but instead of using the square of the coefficients,
absolute values of magnitudesare used Regularization 类似于"岭回归"，但不是使用系数的平方, 使用幅度的绝对值
Chi-squared Automatic Interaction Detection (CHAID) 卡方自动交互检测(CHAID)
Ridge Regression (L2) Method in which the "ordinary least squares" loss (or cost) function is altered by adding a penalty equivalent to square
the of the magnitude of the coefficients 岭回归(L2) 方法，其中"普通最小二乘"损失（或成本）函数是通过添加一个惩罚等效于系数大小的平方
Classification and Regression Tree (CART)
Umbrella term used to refer to classification (predicted outcome is a “class”) and regression trees (predicted outcome is a "number") 用于指代分类（预测结果是"类"）和回归树（预测结果是"数字"）的伞形术语回归树(CART)
Conditional Decision Trees 条件决策树

Given an agent that receives "rewards" by interacting with an environment, the goal of RL is to learn a strategy whereby the agent maximizes future rewards 给定一个通过与环境交互获得"奖励"的代理，RL的目标是学习一种策略，使代理最大化未来的奖励
Basic: uses only a single attribute for splitting Decision Stump 基本：仅使用单个属性进行拆分决策树桩 "政策梯度定理"=预期奖励的导数是政策的日志的奖励和梯度的乘积的期望
"Policy Gradient Theorem" = the derivative of the expected reward is the 强化算法计算策略梯度
The REINFORCE algorithm computes the policy gradient
Policy gradient /REINFORCE expectation of the product of the reward and gradient of the log of the policy 通常用于NLP问题政策梯度/加强
Typically used for NLP problems Iterative Dichotomiser 3 (ID3) Iterative Dichotomiser 3 (ID3)
The basic principle uses gradient ascent to follow policies with the steepest increase in rewards 基本原理使用梯度上升来遵循奖励增加最陡的政策
A binary decision tree learner with linear regression functions at the terminal (leaf) nodes (that can predict continuous numerical attributes) M5 在终端（叶）节点具有线性回归函数的二元决策树学习器（可以预测连续的数值属性） M5
Asynchronous: Unlike DQN, which uses a single agent and a single environment, A3C uses multiple agents with each agent having its own network parameters and a copy of the environment 异步：与使用单个代理和单个环境的DQN不同，A3C使用多个代理，每个代理都有自己的网络参数和环境的副本
See Ensembles... 见合奏。..
Random forests / Random DT Actor-Critic: Predicts both the value function V(s) as well as the optimal policy function,
p (s) 随机森林/随机DT Actor-Critic：预测值函数V（s）以及最优策略函数p（s）
Designed to mitigate "basic" DT methods from overfitting to their training set Decision Trees (DT) 旨在减轻"基本"DT方法从过度拟合到其训练集决策树(DT)
"On-policy" - learns value function for one policy while
following it (i.e., cannot learn value function by following another policy). "On-policy"学习一个策略的价值函数，而遵循它（即不能通过遵循另一个策略来学习价值函数）。
Recursive generation of new tree nodes using subset of prior data Step 1: Induction 使用先前数据子集递归生成新树节点步骤1：感应
Advantage Advantage 异步：与使用单个代理和单个环境的DQN不同，A3C使用多个代理，每个代理都有自己的网络参数
Asynchronous: Unlike DQN, which uses a single agent and a single environment, A3C uses multiple 异步优势演员-评论家(A3C)
(To minimize overfitting) prune the unnecessary splits in developing tree Step 2: Pruning Policy optimization Actor-Critic Asynchronous Advantage Actor-Critic (A3C)
agents with each agent having its own network parameters and a copy of the environment （以尽量减少过度拟合）修剪树中不必要的分裂步骤2：修剪策略优化 Actor-Critic 和环境的副本
(A2C) (A2C)
(Arguably SOTA) Method to mitigate convergence issues that "policy gradient" methods have （可以说是SOTA）缓解"政策梯度"方法所具有的收敛问题的方法
Proximal Policy Optimization (PPO) / OpenAI Strikes a balance between ease of implementation, sample complexity, and ease 近端策略优化(PPO)/OpenAI 在易于实现、样本复杂度和易于调优之间取得平衡，尝试在每个步骤计算一个更
Discriminative of tuning, trying to compute an update at each step that minimizes the cost Discriminative 新，使成本函数最小化，同时确保与先前策略的偏差相对较小
function while ensuring the deviation from the previous policy is relatively small
Parses entire dataset to find k-nearest instances to the new instance 解析整个数据集以查找与新实例最近的k个实例
Updates policies by taking the largest step possible to improve performance, while 通过采取尽可能大的步骤来更新策略以提高性能，同时满足新策略和旧策略允许的接
Trust Region Policy Optimization (TRPO) 信任区域策略优化(TRPO)
satisfying a special constraint on how close the new and old policies are allowed to be 近程度的特殊约束
The objective of a RL agent is to maximize the “expected” reward when following pa policy RL代理的目标是在遵循策略p时最大化"预期"奖励
k-Nearest Neighbors (kNN) k-Nearest Neighbors (kNN) 关键技术是"体验重播"，dqn存储包括状态转换、奖励和操作在内的体验，
Key technique is "Experience Replay" whereby DQN stores experiences 优化以获得更稳定的
Optimized to acheive 深度Q神经网络(Dqn)/DeepMind 这些都是执行Q学习所必需的数据，并制作小批量来更新NNs
Unlike LR, output value is categorical (i.e., binary valued) rather than a numeric value Data points modeled using the standard logistic (i.e., sigmoid) function Logistic Regression (LogR) Deep Q Neural Network (DQN) / DeepMind Q-learning with NNs including state transitions, rewards and actions, which are necessary data 与LR不同，输出值是分类值（即二进制值），而不是数字值使用标准logistic（即sigmoid）函数建模的数据点逻辑回归(LogR) Q-learning with NNs 训练
more stable training to perform Q learning, and makes mini-batches to update NNs
Model-free Model-free
Categorical 51-Atom DQN (c51) 51 represents the use of 51 discrete values to parameterize the value distribution, Z(s,a), w/ the Distributional Bellman equation Categorical 51-Atom DQN (c51) 51表示利用51个离散值对值分布进行参数化，Z(s,a),w/所述分布贝尔曼方程
Value Based / Distributional Reinforcement Learning 基于价值 / 分位数回归的分布强化学习(QR-DQN)

For each state-action pair instead of estimating a single value a distribution of values values in learned 对于每个状态-动作对，而不是估计单个值，在学习中值的分布
Q-learning with Quantile Regression (QR-DQN) Q-learning
Given a set of points of 2 types in N dimensional place, an SVM generates an (N - 1) dimensional hyperplane to separate those points into 2 groups Designed to mitigate both "sparse" and Unlike in DQN (in which a buffer of past experiences is used to stabilize training), HER effectively 给定一组在N维位置的2类型的点，SVM生成一个（N1）维超平面以将这些点分成2组旨在缓解"稀疏"和"二元"奖励"问题以及多目与dqn（使用过去经验的缓冲区来稳定训练）不同，她有效地添加了虚构的数据，通过想象如果情
Hindsight Experience Replay (HER) / OpenAI 后见之明经验重播(她)/OpenAI
"binary" reward" problems and multi-goal RL adds fictitious data, by imagining what would happen had the circumstance been different 标RL 况不同会发生什么
An off-policy algorithm that seeks to find the best action to take given the Off-policy because the Q-learning function learns from actions outside the 寻求在当前状态下找到最佳行动的非策略算法（即，寻求学习使总奖励最关闭策略是因为Q-learning功能从当前策略之外的操作中学习，例如采取随机操作
current state (i.e., seeks to learn a policy that maximizes the total reward) current policy, like taking random actions, and therefore a policy isn’t needed 大化的策略），因此不需要策略
Support Vector Machine (SVM) 支持向量机(SVM)
Classification Deep Deterministic Policy Gradients (DDPG) A model-free off-policy algorithm for learning continous actions Combines ideas from DPG (Deterministic Policy Gradient) and DQN
Classification 深度确定性政策梯度(DDPG) 一种学习连续动作的无模型非策略算法结合Dpg（确定性政策梯度）和DQN的想法
A generative classifier yields the conditional probability of observable X, given target y: P(X |Y=y) 生成分类器产生可观察X的条件概率，给定目标y：P（X/Y=y）
Soft Actor -Critic (SAC) Optimizes a stochastic policy in an off-policy way, forming a bridge between stochastic policy optimization and DDPG-style approaches 软演员-评论家(SAC) 以非政策的方式优化随机政策，在随机政策优化和DDPG式方法之间架起一座桥梁
Explicitly apply Bayes’ Theorem to (typically, classification and regression) problems Reinforcement Learning (RL) Hybrid: combine the 明确地将贝叶斯定理应用于（通常是分类和回归）问题强化学习(RL) 混合：结合Q学习和政
strengths of Q-learning Learns two Q-functions instead of one (hence “twin”) 策梯度的优势学习两个Q函数而不是一个（因此是"孪生"）
Assumes independence between predictors "Simple" application of Bayes' Theorem Twin Delayed Deep 假定预测变量之间的独立性贝叶斯定理的"简单"应用双延迟深度确定性
and policy gradients Mitigates issues with DDPG: e.g., it is frequently brittle with Updates policy (and target networks) less frequently than Q-function 政策
缓解DDPG的问题：例如，对于超参数和其他类型的调整，更新策略（和目标网络）的频率低于Q-function
Deterministic Policy 它经常是脆弱的
Features are vectorised in a binary fashion BernoulliNB respect to hyperparameters and other kinds of tuning 特征以二进制方式进行矢量化 BernoulliNB
Gradients (TD3) Adds noise to the target action, to make it harder for the policy to Gradients (TD3) 将噪声添加到目标操作中，使策略更难利用q函数错误，方法是沿着
exploit Q-function errors by smoothing out Q along changes in action 操作的变化平滑Q
ComplementNB
Variant of MNB that uses uses statistics from the complement of each class to compute the model’s weight Naive Bayes (NB) MNB的变体使用来自每个类的补码的统计来计算模型的权重 lementNB Naive Bayes (NB)
Temporal Difference (TD) Calculates Q-values with respect to changes in the environment over time Combine the sampling of Monte Carlo with thebootstrapping of Dynamic Programming 基本假设是特征遵循正态分布时间差异(TD) 计算环境随时间变化的Q值将蒙特卡洛的采样与动态规划的引导相结合
Underlying assumption is that the features follow a normal distribution GaussianNB GaussianNB
Bayesian Algorithms With policy optimization methods one directly optimizes for the thing one wants
贝叶斯算法
Non-binary version of BernoulliNB MultinomialNB (MNB) 非二进制版本的BernoulliNB MultinomialNB (MNB) 通过策略优化方法，人们可以直接针对自己想要的东西进行优化
Trade-offs Between Policy Optimization and Q-Learning 策略优化和Q-Learning之间的权衡
Designed to mitigate attribute-independence in NB Averaged One-Dependence Estimators (AODE) With Q-learning methods one only indirectly optimizes for agent performance, by training to satisfy a self-consistency equation 旨在减轻NB中的属性独立性 Averaged One-Dependence Estimators (AODE) 使用Q-learning方法，只会通过训练来满足自一致性方程来间接优化代理性能
Model-Based Priors for Model-Free Aims to bridge gap between 每个层中的节点都连接到上一层和下一层中的所有节点设计为使每个层学习整个输入贝叶斯信念网络(BBN)/深度信念网络(DBN) 基于模型的无模型前因旨在弥合无模型和基于模型的R
Nodes in each layer are connected to all the nodes in the previous and subsequent layer Designed so that each layer learns the entire input Bayesian Belief Network (BBN) / Deep Belief Networks (DBN) Learne probabilistic dynamics model and leverages it as a prior for the intertwined model-free optimization 学习概率动力学模型，并将其作为交织无模型优化的先验
Reinforcement Learning (MBMF) model-free and model-based RL 强化学习(MBMF) L之间的差距
Typically used when groups are a priori known (unlike in clustering) Find linear combination of features that characterizes or separates two or more classes of objects or events Linear Discriminant Analysis (LDA) 通常用于先验已知组时（与聚类中不同）查找表征或分隔两个或多个对象或事件类的要素的线性组合线性判别分析(lda)
Model-Based Value Expansion (MBVE) Controls for uncertainty in the model by only allowing imagination to fixed depth Enables wider use of learned dynamics models within a model-free RL algorithm 基于模型的价值扩展(MBVE) 通过只允许想象固定深度来控制模型中的不确定性在无模型RL算法中更广泛地使用学习的动力学模型
The generative network's goal is to increase One network generates candidates, Learn the model 生成网络的目标是提高判别网络的错误率一个网络生成候选者，而网络对其学习模型
Given a training set, GANS learn to generate new data with the same statistics as the training set World Models E.g., agents learn from own “dreams” generated by Variable Auto-encoders 给定一个训练集，我们学习生成与训练集具有相同统计信息的新数据世界模型例如，代理从可变自动编码器生成的自己的"梦想"中学习
error rate of the discriminative network while the network evaluates them 进行评估
The most basic approach never explicitly represents the Each time an agent observes environment, it computes a plan which 最基本的方法从不明确表示策略，而是使用纯粹的计划每次代理观察环境时，它都会计算一个相对于模型最优的计划，其中
A generative model G captures data distribution, and a discriminative model D estimates probability that a sample came from the training data rather than G Original GAN (Goodfellow, et al.) policy, and instead, uses pure planning techniques like is optimal with respect to the model, where the plan describes all
生成模型G捕获数据分布，判别模型D估计样本来自训练数据而不是G的概率原GAN(Goodfellow，et al.)
技术，如模型预测控制（mpc）来选择操作该计划描述了在当前之后接管某个固定时间窗口的所有操作
Learns a hierarchy of representations from object parts to scenes in both the G and D Deep Convolutional Generative Adversarial Network (DCGAN) Generative Model-Based model-predictive control (MPC) to select actions actions to take over some fixed window of time after the present
学习g和D中从对象部分到场景的表示层次结构深度卷积生成对抗网络(DCGAN) Generative Model-Based
Agent uses a planning algorithm (like MTCS) in model to generate E.g., AlphaZero 代理在模型中使用计划算法（如MTCS），通过从其当前策略中采 E.g., AlphaZero
Provides a new convergence measure, fast and stable training and high visual quality Boundary Equilibrium Generative Adversarial Networks (BEGAN) Given the model 提供了新的收敛度量，训练快速稳定，视觉质量高边界平衡生成对抗网络（开始）给定模型
candidate actions for the plan by sampling from its current policy 样来生成计划的候选操作例如，退出（对于十六进制棋盘游戏）
E.g., ExIt (for Hex board game)
Mitivated by "network embedding" problem = to learn a low-dimensional representation for each node of a 通过"网络嵌入"问题=学习网络每个节点的低维表示，以有利于下游任务，如节点分类、链路预测和网络可
network to benefit downstream tasks, such as node classification, link prediction, and network visualization Learns to interpret predictions from a learned environment model to construct implicit plans in 视化学习从已学习的环境模型中解释预测，以任意方式构造隐式计划，将预测用作深层策略网络中
Proximity Generative Adversarial Network (ProGAN) Imagination-Augmented arbitrary ways, by using the predictions as additional context in deep policy networks 邻近生成对抗网络(ProGAN) Imagination-Augmented 结合了无模型和基于模型的附加上下文
Need to discover proximity between different nodes to learn good node representations Generative Adversarial Combines model-free and 需要发现不同节点之间的接近度，以学习良好的节点表示生成性对抗
Agents (I2A) model-based aspects Embeds the planning procedure directly into a policy as a subroutine (so that complete plans become side Agents (I2A) 的方面将规划过程作为子例程直接嵌入到策略中（以便完整的计划成为策略的侧面信息），同时使用任何标准的无
ProGAN designed to generate proximities (that benefits network embedding) Networks (GANs) information for the policy) while training the output of the policy with any standard model-free algorithm ProGAN旨在生成近似值（有利于网络嵌入） Networks (GANs) 模型算法训练策略的输出
Find that applying orthogonal regularization to G renders it amenable to a "truncation trick," allowing 发现对G应用正交正则化使其适合于"截断技巧"，通过减少G输入的方差，允许对样本保真度和多样性
Achieved SOTA (c.2019) in class-conditional image synthesis BigGAN RL typically used to "solve" problems for which decision making is sequential, but the goal is long-term (e.g., game playing, robotics, resource management, or logistics) 在类条件图像合成中实现SOTA(c.2019) BigGAN RL通常用于"解决"决策是顺序的问题，但目标是长期的（例如，游戏，机器人，资源管理或物流）
fine control over trade-off between sample fidelity and variety by reducing the variance of G's input 之间的权衡进行精细控制
Numerous algorithms; see GAN Zoo 众多算法;见GAN动物园
(Introduced after "Original GAN") Method generates anindependent sample via a single feedforward pass through a multilayer perceptron （在"原始GAN"之后引入）方法通过通过多层感知器的单个前馈通道生成独立样本
Generative Moment Matching Networks (GMMs) 生成矩匹配网络(GMMs)
Maximum Mean Discrepancy (MMD): match all orders of statistics between a dataset and samples from model; train via backpropagation 最大平均差异（MMD）：匹配数据集和模型样本之间的所有统计顺序；通过反向传播训练
Figures Figures
Figure 1. Mechanisms for civilian harm .............................................................................................. 14 Figure 1. 平民伤害机制。............................................................................................. 14

Figure 2. Civilian protection life cycle .................................................................................................. 17 Figure 2. 平民保护生命周期。................................................................................................. 17
Figure 3. CPLC-AI matrix ........................................................................................................................... 25 Figure 3. CPLC-AI matrix ........................................................................................................................... 25
Figure 4. Notional schematic illustrating key elements of the civilian harm AI Figure 4. 说明平民伤害AI问题关键要素的名义示意图。......................................................................................
problem ......................................................................................................................................... 45 .................................................. 45
Figure 5. Schematic adding some inherent uncertainty to the notionally “perfect” Figure 5. 示意性地向图4中的概念上"完美"设置添加一些固有的不确定性。...............................
setting in Figure 4 ...................................................................................................................... 46 ...................................................................................... 46
Figure 6. Schematic adding a second layer of ambiguity or uncertainty to Figure 4 ....... 47 Figure 6. 示意性地向图4添加模糊性或不确定性的第二层。...... 47
Figure 7. CIVCAS Cause-Mitigations-AI-applications (CIVCAS-AI) Matrix ............................ 48 Figure 7. CIVCAS Cause-Mitigations-AI-applications (CIVCAS-AI) Matrix ............................ 48
Figure 8. The Five Tribes of AI Functions ........................................................................................... 54 Figure 8. AI功能的五个部落。.......................................................................................... 54
Figure 9. High level taxonomy of machine learning methods .................................................... 57 Figure 9. 机器学习方法的高级分类学。................................................... 57
Figure 10. Full taxonomy of machine learning methods ................................................................. 58 Figure 10. 机器学习方法的全分类法。................................................................ 58
59 59
Abbreviations Abbreviations
ABM agent-based modeling ABM agent-based modeling

AI artificial intelligence AI 人工智能
ATR-MCAS Aided Threat Recognition from Mobile Cooperation and Autonomous ATR-MCAS 来自移动合作和自主的辅助威胁识别
Sensors Sensors
C2 command and control C2 指挥及控制
C2IMERA Command and Control Incident Management Emergency Response C2IMERA 指挥和控制事件管理应急响应
Application Application
CATE collaborative AI at the tactical edge CATE 战术边缘的协作人工智能
CCW Convention for Certain Conventional Weapons CCW 某些常规武器公约
COMPASS Collection and Monitoring via Planning for Active Situational COMPASS 通过积极情景规划进行收集和监测
Scenarios Scenarios
CPLC civilian protection life cycle CPLC 平民保护生命周期
CST complex system theory CST 复杂系统理论
CV Computer Vision CV 计算机视觉
DARPA Defense Advanced Research Projects Agency DARPA 国防高级研究计划局
DL Deep learning DL 深入学习
DOD Department of Defense DOD 国防部
IHL international humanitarian law IHL 国际人道主义法
ISAF International Security Assistance Force ISAF 国际安全援助部队
KR&R Knowledge Representation and Reasoning KR&R 知识表示与推理
LAWS lethal autonomous weapon systems LAWS 致命的自主武器系统
ML machine learning ML 机器学习
NLP Natural Language Processing NLP 自然语言处理
NSCAI National Security Commission on Artificial Intelligence NSCAI 人工智能国家安全委员会
OODA observe, orient, decide, act OODA 观察，定向，决定，行动
P&S Planning and Scheduling P&S 计划及时间表
SA situational awareness SA 态势感知
UAS unmanned aerial system UAS 无人飞行系统
UAV unmanned aerial vehicle UAV 无人飞行器
UK United Kingdom UK 联合王国
UN United Nations UN 联合国
USV unmanned surface vessel USV 无人水面舰艇
60 60
References References
Abdelhamid, Abdelaziz, Hamzah Alsayadi, Islam Hegazy, and Zaki Taha Fayed. “End-to-End Arabic Abdelhamid，Abdelaziz，Hamzah Alsayadi，Islam Hegazy和Zaki Taha Fayed。 "端到端阿拉伯语
Speech Recognition: A Review.” In 19th Conference on Language Engineering (ESOLEC’19), Sept. 语音识别：回顾。"在第19届语言工程会议（ESOLEC'19），9月。 2020.
2020.
Ackerman Evan, and Michael Koziol. “In the Air with Zipline’s Medical Delivery Drones.” IEEE Spectrum 阿克曼*埃文和迈克尔*科齐奥尔。 "在空中与Zipline的医疗送货无人机。"IEEE频谱（Apr。 30, 2019)
(Apr. 30, 2019). https://spectrum.ieee.org/in-the-air-with-ziplines-medical-delivery- . https://spectrum.ieee.org/in-the-air-with-ziplines-medical-deliverydrones/particle-1 ...
drones/particle-1.
“Adaptable Command and Control System Allowed Georgia Air Force Base to Track and Prepare for "适应性强的指挥和控制系统使格鲁吉亚空军基地能够跟踪和准备
Hurricane Dorian.” Leidos. Oct. 1, 2019. https://www.leidos.com/insights/adaptable-command- 飓风多利安。"莱多斯。十月。 1, 2019. https://www.leidos.com/insights/adaptable-comma
and-control-system-allowed-georgia-air-force-base-track-and-prepare. nd 和-控制-系统-允许-格鲁吉亚-空军-基地-轨道-和-准备。
Adewole, Sodiq, Erfaneh Gharavi, Benjamin Shpringer, Martin Bolger, Vaibhav Sharma, Sund Ming Adewole,Sodiq,Erfaneh Gharavi,Benjamin Shpringer,Martin Bolger,Vaibhav Sharma,Sund Ming
Yang, and Donald E. Brown. “Dialogue-Based Simulation for Cultural Awareness Training.” Feb. 杨和唐纳德*E*布朗。 "基于对话的模拟文化意识培训。"二月。 1, 2020. arXiv：2002.00223
1, 2020. arXiv:2002.00223v1 [cs.CY]. v1[cs.CY]。
Aggarwal, Ravi, Viknesh Sounderajah, Guy Martin, Daniel S. W. Ting, Alan Karthikesalingam, Dominic Aggarwal,Ravi,Viknesh Sounderajah,Guy Martin,Daniel S.W.Ting,Alan Karthikesalingam,Dominic
King, Hutan Ashrafian, et al. “Diagnostic Accuracy of Deep Learning in Medical Imaging: A King，Hutan Ashrafian，et al。 "医学成像中深度学习的诊断准确性：A
Systematic Review and Meta-Analysis.” npj Digital Medicine 4, article 65 (Apr. 7, 2021). 系统回顾和荟萃分析。"npj数字医学4，第65条（Apr。 7, 2021). https://doi.org/10.
https://doi.org/10.1038/s41746-021-00438-z. 1038/s41746-021-00438-z ...
Allesina, Stefano, and Mercedes Pascual. “Googling Food Webs: Can an Eigenvector Measure Species' Allesina，Stefano和Mercedes Pascual。 "谷歌搜索食物网：特征向量可以测量物种吗'
Importance for Coextinctions?” PLoS Computational Biology 5, no. 9 (2009). 共存的重要性？"PLoS计算生物学5，no.9（2009）。
Amazon. “In the News: How Amazon Is Using Machine Learning to Eliminate 915,000 Tons of 亚马逊。 "在新闻中：亚马逊如何使用机器学习来消除915,000吨的
Packaging.” Jan. 29, 2021. https://aws.amazon.com/blogs/industries/how-amazon-is-using- Packaging.” Jan. 29, 2021. https://aws.amazon.com/blogs/industries/how-amazon-is-using-
machine-learning-to-eliminate-915000-tons-of-packaging/. machine-learning-to-eliminate-915000-tons-of-packaging/.
Australia, Canada, Japan, South Korea, the United Kingdom, and the United States. Discussion Paper – 澳大利亚，加拿大，日本，韩国，英国和美国。讨论文件 –
Building on Chile’s Proposed Four Elements of Further Work for the Convention on Certain 在智利提出的《关于某些问题的公约》进一步工作的四个要素的基础上
Conventional Weapons (CCW) Group of Governmental Experts (GGE) on Emerging Technologies in 常规武器（常规武器公约）致命自主武器系统（法律）领域新兴技术政府专家组。提交给联
the Area of Lethal Autonomous Weapons Systems (LAWS). Submitted to the UN CCW June 2021. 合国CCW六月2021。
Badia, Adria, Bilal Piot, Steven Kapturowski, Pablo Sprechmann, Alex Vitvitskyi, Daniel Guo, and Badia，Adria，Bilal Piot，Steven Kapturowski，Pablo Sprechmann，Alex Vitvitskyi，Daniel Guo和
Charles Blundell. “Agent57: Outperforming the Atari Human Benchmark.” Mar. 30, 2020. 查尔斯*布伦德尔。 "Agent57：表现优于Atari人类基准。"玛尔． 30, 2020. arXiv：2
arXiv:2003.13350. 003.13350。
Baker, Chris, Sarvapali Ramchurn, W.T. Luke Teacy, and Nicholas R. Jennings. “Planning Search and Baker，Chris，Sarvapali Ramchurn，W.T.Luke Teacy和Nicholas R.Jennings。 "规划搜寻及
Rescue Missions for UAV Teams.” In Proceedings of the Twenty-second European Conference on 无人机团队的救援任务。"在第二十二届欧洲会议记录中
Artificial Intelligence, Aug. 2016. 人工智能，8月8日。 2016.
Barker, Alex. “Giving Precision Munitions ‘Eyes’ and a ‘Brain’: The State of PLA Research on Military 巴克，亚历克斯。 "赋予精确弹药'眼睛'和'大脑'：解放军军事研究的现状
Target Recognition.” China Brief 21, no. 13 (July 2, 2021). 目标识别。"中国简报21，第13期（2021年7月2日）。 https://jamestown.org/program/givi
https://jamestown.org/program/giving-precision-munitions-eyes-and-a-brain-the-state-of-pla- ng-precision-munitions-eyes-and-a-brain-the-state-of-pla research-on-military-target-
research-on-military-target-recognition/. recognition/。
61 61
Barlos, Fotis, Anna Skinner, Richard Peeke, Robert Mohan, Andjelka Kelic, Walter Beyeler, Rossitza Barlos,Fotis,Anna Skinner,Richard Peeke,Robert Mohan,Andjelka Kelic,Walter Beyeler,Rossitza
Homan, et al. Collection and Monitoring via Planning for Active Situational Scenarios (COMPASS): 霍曼等人。通过规划活动情境场景（COMPASS）进行收集和监控:
Strategic Multi-Layer Assessment Report. Sandia National Laboratories. SAND2020-0136R. Jan. 战略多层评估报告。桑迪亚国家实验室。桑德2020-0136r.1月 2020. https://www.osti.g
2020. https://www.osti.gov/servlets/purl/1592839. ov/servlets/purl/1592839 ...
Biggs, Adam, Matthew Cain, and Stephen Mitroff. “Cognitive Training Can Reduce Civilian Casualties in 比格斯，亚当，马修该隐和斯蒂芬米特罗夫。 "认知训练可以在模拟射击环境中减少平民伤亡。"心
a Simulated Shooting Environment.” Psychological Science (July 13, 2015). 理科学（2015年7月13日）。
Blasch, Erik, Tien Pham, Chee-Yee Chong, Wolfgang Koch, Henry Leung, Dave Braines, and Tarek Blasch，Erik，Tien Pham，Chee-Yee Chong，Wolfgang Koch，Henry Leung，Dave Braines和Tarek
Abdelzaher. “Machine Learning/Artificial Intelligence for Sensor Data Fusion–Opportunities and 阿卜杜勒扎赫。 "用于传感器数据融合的机器学习/人工智能-机会和
Challenges.” IEEE Aerospace and Electronic Systems Magazine 36, no. 7 (July 2021). 挑战。"IEEE航空航天和电子系统杂志36，第7期（2021年7月）。
Blumenstein, Richard. “Kessel Run’s C2IMERA Used During Afghan Evacuation.” Wright-Patterson AFB. 布卢门斯坦，理查德。 "Kessel Run的C2imera在阿富汗撤离期间使用。"赖特-帕特森空军基地。
Sept. 23, 2021. https://www.wpafb.af.mil/News/Article-Display/Article/2786182/kessel-runs- Sept. 23, 2021. https://www.wpafb.af.mil/News/Article-Display/Article/2786182/kessel-runs-
c2imera-used-during-afghan-evacuation/. c2imera-used-during-afghan-evacuation/.
Brown, Noam, and Tuomas Sandholm. “Superhuman AI for Multiplayer Poker.” Science 365 (Aug. 30, 布朗，诺姆和图奥马斯桑德霍尔姆。 "多人扑克的超人AI。"科学365（8月。 30, 2019).
2019).
Burns, Lawrence D. Autonomy: The Quest to Build the Driverless Car—And How It Will Reshape Our Burns，Lawrence D.Autonomy：构建无人驾驶汽车的追求-以及它将如何重塑我们的
World. New York: Ecco Press, 2018. 世界。缧√＃陨上□ 辅缧
“Can a Pandemic Drone Help Stop the Spread of COVID-19?” Draganfly website. July 3, 2020. "大流行无人机可以帮助阻止COVID-19的传播吗？"Draganfly网站。 2020年7月3日。
https://draganfly.com/news/can-a-pandemic-drone-help-stop-the-spread-of-covid-19/. https://draganfly.com/news/can-a-pandemic-drone-help-stop-the-spread-of-covid-19/.
Carvalho, L.E., and A. von Wangenheim. “3D Object Recognition and Classification: A Systematic Carvalho，L.E.和a.von Wangenheim。 "三维物体识别与分类：系统化
Literature Review.” Pattern Analysis and Applications 22 (Feb. 2019). 文献综述."模式分析与应用22（2月22日）。 2019).
Chandra Yogesh, and Antoreep Jana. “Sentiment Analysis Using Machine Learning and Deep Learning.” Chandra Yogesh和Antoreep Jana。 "使用机器学习和深度学习进行情感分析。”
In 7th International Conference on Computing for Sustainable Global Development, New Delhi, 第七届计算促进全球可持续发展国际会议，新德里,
India, Mar. 12-14, 2020, IEEE. 印度，3月。 12-14,2020,IEEE.
Chen, Chao, Weidong Bao, Tong Men, Xiaomin Zhu, Ji Wang, and Rui Wang. “NECTAR-An Agent-Based 陈、赵、卫东宝、童门、朱小敏、吉王和瑞王. "花蜜-基于代理
Dynamic Task Allocation Algorithm in the UAV Swarm.” Complexity 2020 (Sept. 2020). 无人机群中的动态任务分配算法。"复杂性2020（9月）。 2020).
Chen, Liang-Chu, Chia-Meng Lee, and Mu-Yen Chen. “Exploration of Social Media for Sentiment Analysis 陈良珠、李家孟和陈慕彦. "社交媒体用于情感分析的探索
Using Deep Learning.” Soft Computing 24 (2020). 使用深度学习。"软计算24（2020）。
Chung, Timothy H. “OFFensive Swarm-Enabled Tactics (OFFSET).” Mar. 2021. Chung，Timothy H."进攻群启用战术（偏移）。"玛尔． 2021.
https://apps.dtic.mil/sti/pdfs/AD1125864.pdf. https://apps.dtic.mil/sti/pdfs/AD1125864.pdf.
Cody, Jason, Karina Roundtree, and Julie Adams. “Human-Collective Collaborative Target Selection.” 科迪，杰森，卡琳娜*罗德特里和朱莉*亚当斯。 "人类-集体协作目标选择。”
ACM Transactions on Human-Robot Interaction 10, no. 2 (Mar. 2021). ACM人机交互交易10,第2期(3月). 2021).
Cook, Christina. “Designing a Virtual Embedded Scenario-Based Military Simulation Training Program 库克，克里斯蒂娜。 "设计基于虚拟嵌入式场景的军事模拟训练程序
Using Educational and Design Instructional Strategies.” Doctoral thesis, University of Central 使用教育和设计教学策略."中央大学博士论文
Florida, 2018. Florida, 2018.
Cox, Mathew. “Army’s Next Infantry Weapon Could Have Facial-Recognition Technology.” Real Clear 考克斯，马修。 "陆军的下一个步兵武器可能有面部识别技术。"真的很清楚
Defense. June 3, 2019. 辩护。 2019年6月3日。
Cristianini, Nello. An Introduction to Support Vector Machines and Other Kernel-Based Learning 克里斯蒂亚尼，内洛。支持向量机和其他基于内核的学习简介
Methods. New York: Cambridge University Press, 2000. 方法。呐
62 62
“DARPA Subterranean Challenge Announces Systems Competition Teams for Final Event.” DARPA "DARPA Subterranean Challenge宣布系统竞赛团队参加决赛。"DARPA网站。 2021年5月3日。
website. May 3, 2021. https://www.darpa.mil/news-events/2021-05-03. https://www.darpa.mil/news-events/2021-05-03...
“DARPA Tiles Together a Vision of Mosaic Warfare.” DARPA website. Sept. 24, 2018. "DARPA将马赛克战争的愿景拼凑在一起。"DARPA网站。 "九月。 24, 2018.
https://www.darpa.mil/work-with-us/darpa-tiles-together-a-vision-of-mosiac-warfare. https://www.darpa.mil/work-with-us/darpa-tiles-together-a-vision-of-mosiac-warfare.
DeBerry, William, Richard Dill, Kenneth Hopkinson, Douglas Hodson, and Michael Gramaila. “The DeBerry，William，Richard Dill，Kenneth Hopkinson，Douglas Hodson和Michael Gramaila。 "的
Wargame Commodity Course of Action Automated Analysis Method.” The Journal of Defense Wargame商品行动过程自动化分析方法。"国防杂志
Modeling and Simulation: Applications, Methodology, Technology (July 19, 2021). 建模与仿真：应用，方法，技术（2021年7月19日）。
Dellerman, Dominik, Adrian Calma, Nikolaus Lipusch, Thorsten Weber, Sascha Weigel, and Philipp Dellerman，Dominik，Adrian Calma，Nikolaus Lipusch，Thorsten Weber，Sascha Weigel和Philipp
Ebel. “The Future of Human-AI Collaboration: A Taxonomy of Design Knowledge for Hybrid 埃贝尔人类-人工智能协作的未来：混合设计知识的分类
Intelligence Systems.” May 7, 2021. arXiv:2105.03354 [cs.AI]. 情报系统。"2021年5月7日。 arXiv：2105.03354[cs.AI]。
Department of Defense. Foreign Internal Defense. Joint Publication 3-22. Aug. 17, 2018. 国防部。外国内部防御。联合出版物3-22。八月。 17, 2018. https://www.jcs.mil/
https://www.jcs.mil/Portals/36/Documents/Doctrine/pubs/jp3_22.pdf. Portals/36/Documents/Doctrine/pubs/jp3_22.pdf...
Department of Defense. Joint Staff civilian harm Review. Apr. 17, 2018. 国防部。联合工作人员平民伤害审查。四月。 17, 2018.
https://www.justsecurity.org/wp-content/uploads/2019/02/Civ-Cas-Study-Redacted-just- https://www.justsecurity.org/wp-content/uploads/2019/02/Civ-Cas-Study-Redacted-just-
security.pdf. security.pdf.
Department of Defense. “Mission Initiatives: Joint Warfighting Operations.” JAIC website. 国防部。 "任务倡议：联合作战行动。"JAIC网站。
https://www.ai.mil/mi_joint_warfighting_operations.html. https://www.ai.mil/mi_joint_warfighting_operations.html.
Department of Defense. “Mission Initiatives: Threat Reduction and Protection.” JAIC website. 国防部。 "特派团倡议：减少威胁和保护。"JAIC网站。 https://www.ai.mil/mi_threat_red
https://www.ai.mil/mi_threat_reduction_and_protection.html. uction_and_protection.html...
De Simone, Sergio. “Ushahidi and the Power of Crowdsourcing.” InfoQ. June 27, 2015. 德西蒙尼，塞尔吉奥。 "Ushahidi和众包的力量。"InfoQ。 2015年6月27日。
De Spiegeleire, Stephan, Matthijs Maas, and Tim Sweijs. Artificial Intelligence and the Future of Defense: De Spiegeleire，Stephan，Matthijs Maas和Tim Sweijs。人工智能与国防的未来:
Strategic Implications for Small and Medium-Sized Force Providers. The Hague Centre for 对中小型部队提供者的战略影响。海牙中心
Strategic Studies. 2017. https://www.jstor.org/stable/resrep12564. 战略研究。 2017. https://www.jstor.org/stable/resrep12564...
Domingos, Pedro. The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake 多明戈斯，佩德罗。主算法：终极学习机器的追求将如何重塑
Our World. New York: Basic Books, 2015. 我们的世界。纽约：基本书籍，2015年。
Edmonds, Jeffrey, Samuel Bendett, Anya Fink, Mary Chestnut, Dmitry Gorenburg, Michael Kofman, 埃德蒙兹，杰弗里，塞缪尔本德特，安雅芬克，玛丽栗子，德米特里戈伦堡，迈克尔科夫曼,
Kasey Stricklin, et al. Artificial Intelligence and Autonomy in Russia. CNA. DRM-2021-U-029303- Kasey Stricklin等人。人工智能和俄罗斯的自治。 CNA。 DRM-2021-U-029303
Final. May 2021. 最后。 2021年5月。
Ekelhof, Merel A. C. “Lifting the Fog of Targeting: ‘Autonomous Weapons’ and Human Control Through 埃克尔霍夫，梅雷尔A.C."解除目标的迷雾：'自主武器'和通过军事目标镜头的人类控制。"海战学院
the Lens of Military Targeting.” Naval War College Review 71, no. 3 (2018). 评论71，第3期（2018）。
El-Jawad, Mohammed, Rania Hodlod, and Yasser Omar. “Sentiment Analysis of Social Media Networks El-Jawad，Mohammed，Rania Hodlod和Yasser Omar。 "社交媒体网络的情绪分析
Using Machine Learning.” In 14th International Computer Engineering Conference, Cairo, Egypt, 使用机器学习。"在第14届国际计算机工程会议，开罗，埃及,
Dec. 29-30, 2018, IEEE. Dec. 29-30, 2018, IEEE.
Elliot, Joshua. “Artificial Social Intelligence for Successful Teams (ASIST).” DARPA website. Accessed 艾略特，约书亚。 "成功团队的人工智能（ASIST）。"DARPA网站。已访问
Sept. 20, 2021. https://www.darpa.mil/program/artificial-social-intelligence-for-successful- "九月。 20, 2021. https://www.darpa.mil/program/artificial-social-intelligence-for-succ
teams. essful 团队。
Elzayady, Hossam, Khaled Badran, and Gouda Salama. “Arabic Opinion Mining Using Combined CNN - Elzayady，Hossam，Khaled Badran和Gouda Salama。 "使用联合CNN的阿拉伯语意见挖掘
LSTM Models.” International Journal of Intelligent Systems and Applications 12, no. 4 (Aug. 2020). LSTM模型。"国际智能系统与应用杂志12，第4期（8月）。 2020).
63 63
Encyclopedia Britannica. s.v. “artificial intelligence.” Accessed Sept. 20, 2021. 大英百科全书。 s.v."人工智能。"访问9月。 20, 2021.
https://www.britannica.com/technology/artificial-intelligence. https://www.britannica.com/technology/artificial-intelligence.
Fan, Wenhui, Peiyu Chen, Daiming Shi, Xudong Guo, and Li Kou. “Multi-Agent Modeling and Simulation 范、文辉、陈佩玉、石大明、郭旭东、李口。 "AI时代的多智能体建模和仿真。"清华科技26,no.5(O
in the AI Age.” Tsinghua Science and Technology 26, no. 5 (Oct. 2021). ct. 2021).
Ferraris, Patrick. “Aided Detection on the Future Battlefield.” Defense Visual Information Distribution 法拉利，帕特里克。 "未来战场上的辅助探测。"国防视觉信息分发
Service. Jan. 24, 2020. https://www.dvidshub.net/news/360225/aided-detection-future- 服务。简。 24, 2020. https://www.dvidshub.net/news/360225/aided-detection-fut
battlefield. ure 战场。
Forrester, Bruce, and Kees den Hollander. “The Role of Social Media in the Intelligence Cycle.” In Next- Forrester，Bruce和Kees den Hollander。 "社交媒体在智能循环中的作用。"在接下来
Generation Analyst IV, Proceedings Vol. 9851. May 12, 2016. Generation Analyst IV，Proceedings Vol。 9851. 2016年5月12日。
Franceshelli, Giorgio, and Mirco Mosolesi, “Creativity and Machine Learning: A Survey.” Apr. 20, 2021. Franceshelli，Giorgio和Mirco Mosolesi，"创造力和机器学习：一项调查。"四月。 20, 2021. arXiv
arXiv:2104.02726v2. ：2104.02726v2。
GAIA: A Fine-Grained Multimedia Knowledge Extraction System. Blender Lab, UIUC. Video. GAIA：一种细粒度的多媒体知识提取系统。搅拌机实验室，UIUC。视频。
http://blender.cs.illinois.edu/software/gaia-ie/gaia.mp4. http://blender.cs.illinois.edu/software/gaia-ie/gaia.mp4.
GDELT homepage. Accessed Sept. 20, 2021. https://www.gdeltproject.org/. GDELT主页。访问9月。 20, 2021. https://www.gdeltproject.org/。
Geluvaraj, B., P. Satwik, and T. Kumar. “The Future of Cybersecurity: Major Role of Artificial Geluvaraj，B.，P.Satwik和T.Kumar。 "网络安全的未来：人工的主要作用
Intelligence, Machine Learning, and Deep Learning in Cyberspace.” S. Smys, R. Bestak, J.Z. Chen, 网络空间中的智能、机器学习和深入学习."S.Smys，R.Bestak，J.Z.Chen和I.Kotuliak，eds。
and I. Kotuliak, eds. In International Conference on Computer Networks and Communication 在计算机网络和通信技术国际会议上. 新加坡：Springer，2019。
Technologies. Singapore: Springer, 2019.
Gettinger, Dan. The Drone Databook. The Center for the Study of the Drone at Bard College. Sept. 2019. 葛丁格，丹。无人机数据库。巴德学院无人机研究中心。 "九月。 2019.
https://dronecenter.bard.edu/files/2019/10/CSD-Drone-Databook-Web.pdf. https://dronecenter.bard.edu/files/2019/10/CSD-Drone-Databook-Web.pdf.
Ghataoura, Darminder, and Sam Ogbannaya. “Application of Image Captioning and Retrieval to Support Ghataoura，Darminder和Sam Ogbannaya。 "应用图像字幕和检索来支持
Military Decision Making.” In International Conference on Military Communication and 军事决策。"在国际军事通信会议和
Information Systems (ICMCIS), the Hague, Netherlands, May 4-5, 2021, IEEE. 信息系统（ICMCIS），荷兰海牙，2021年5月4-5日，IEEE。
Gregg, Aaron. “In a First, Air Force Uses AI on Military Jet.” Washington Post. Dec. 16, 2020. 格雷格，艾伦。 "首先，空军在军用喷气式飞机上使用人工智能。"华盛顿邮报。十二月。 16, 2020.
https://www.washingtonpost.com/business/2020/12/16/air-force-artificial-intelligence/. https://www.washingtonpost.com/business/2020/12/16/air-force-artificial-intelligence/.
Han, Ji, Serhad Sarica, Feng Shi, and Jianxi Luo. “Semantic Networks for Engineering Design: State of the 韩、纪、瑟哈德*萨里卡、冯世和简西*罗。 "工程设计的语义网络：
Art and Future Directions.” Journal of Mechanical Design 144, no. 2 (Sept. 9, 2021). 术和未来方向。"机械设计杂志144，第2期（9月）。 9, 2021).
Hara, Kotaro, and Jon E. Froehlich, “Characterizing and Visualizing the Physical World Accessibility at Hara，Kotaro和Jon E.Froehlich，"表征和可视化物理世界的可访问性
Scale Using Crowdsourcing, Computer Vision, and Machine Learning.” ACM SIGACCESS 使用众包、计算机视觉和机器学习进行规模化。"ACM SIGACCESS
Accessibility and Computing 113 (Oct. 2015). 可访问性和计算113（Oct。 2015).
Hassoun, Mohamad. Fundamentals of Artificial Neural Networks. Cambridge, MA: MIT Press, 2003. 哈桑，穆罕默德。人工神经网络的基本原理。雁抚蜷医帑九＆缧
He, Kaiming, Xiangu Zhang, Shaoqing Ren, and Jian Sun. “Delving Deep into Rectifiers: Surpassing 何、开明、仙姑张、少卿任、建孙。 "深入研究整流器：超越
Human-Level Performance on ImageNet Classification.” Feb. 6, 2015. arXiv:1502.01852v1 在ImageNet分类上的人级性能。"二月。 6, 2015. arXiv：1502.01852v1[cs.CV]。
[cs.CV].
Hedberg, Sara Reese. “DART: Revolutionizing Logistics Planning.” IEEE Intelligent Systems 17 (May 赫德伯格萨拉*里斯 "DART：彻底改变物流规划。"IEEE智能系统17（2002年5月）。
2002).
Host, Pat. “Deep Learning Analytics Develops DARPA Deep Machine Learning Prototype.” Defense 主持人，帕特。 "深度学习分析开发DARPA深度机器学习原型。"防御
Daily. May 11, 2016. https://www.defensedaily.com/deep-learning-analytics-develops-darpa- 每天。 2016年5月11日。 https://www.defensedaily.com/deep-learning-analytics-develo
deep-machine-learning-prototype/advanced-transformational-technology/. ps-darpa deep-machine-learning-prototype/advanced-transformational-technology/。
64 64
Hou, Shenguluan, and Ruqian Lu. “Knowledge-Guided Unsupervised Rhetorical Parsing for Text 侯，圣古銮，如前路。 "文本的知识引导无监督修辞解析
Summarization.” Information Systems 94 (Dec. 2020). 总结。"信息系统94（Dec。 2020).
IARPA. Request for Information. IARPA-RFI-19-06. Mar. 26, 2019. https://www.iarpa.gov/engage-with- IARPA。索取资料。 IARPA-RFI-19-06。玛尔。 26, 2019. https://www.iarpa.gov/engage-withus/
us/rfi?scroll_position=448.9639892578125&rfi_filter=past. rfi?scroll_position=448.9639892578125&rfi_filter=past...
Ilachinski, Andrew. AI, Robots, and Swarms. CNA. DRM-2017-U-014796-Final. Jan. 2017. 伊拉钦斯基，安德鲁。人工智能，机器人和成群。 CNA。 DRM-2017-U-014796-最终. 简。 2017.
Ilachinski, Andrew. Applications of Social Media to Military Operations: Overview and Assessment. CNA. 伊拉钦斯基，安德鲁。社交媒体在军事行动中的应用：概述和评估。 CNA。
DME-2013-U-005368-Final. July 2013. DME-2013-U-005368-Final. July 2013.
Ilachinski, Andrew. Artificial Intelligence: Emerging Themes, Issues, and Narratives. CNA. DOP-2020-U- 伊拉钦斯基，安德鲁。人工智能：新兴主题，问题和叙述。 CNA。 DOP-2020-U028073-决赛.
028073-Final. Oct. 2020. 十月。 2020.
International Committee of the Red Cross. International Humanitarian Law: Answers to Your Questions. 红十字国际委员会。国际人道主义法：回答你的问题。
June 9, 2020. June 9, 2020.
Ish, Daniel, Jared Ettinger, and Christoper Ferris. Evaluating the Effectiveness of Artificial Intelligence Ish，Daniel，Jared Ettinger和Christoper Ferris。评估人工智能的有效性
Systems in Intelligence Analysis. RAND Corporation. RR-A464-1. 2021. 情报分析系统。兰德公司。 RR-A464-1。 2021.
Jadhav, Nagesh, and Rekha Sugandhi. “Survey on Human Behavior Recognition Using Affective Jadhav，Nagesh和Rekha Sugandhi。 "利用情感识别人类行为的调查"
Computing.” In 2018 IEEE Global Conference on Wireless Computing and Networking (GCWCN), 计算。"2018年IEEE无线计算与网络全球会议(GCWCN),
Lonavala, India, Nov. 23-24, 2018, IEEE. Lonavala，印度，11月11日。 23-24,2018,IEEE.
Janai, Joel, Fatma Guney, Aseem Behl, and Andreas Geiger. “Computer Vision for Autonomous Vehicles: Janai，Joel，Fatma Guney，Aseem Behl和Andreas Geiger。 "自动驾驶汽车的计算机视觉:
Problems, Datasets and State of the Art.” Mar. 17, 2021. arXiv:1704.05519v3 [cs.CV]. 问题，数据集和艺术状态。"玛尔． 17, 2021. arXiv：1704.05519v3[cs.CV]。
Jensen, Benjamin, Scott Cuomo, and Chris Whyte. “Wargaming with Athena: How to Make Militaries Jensen，Benjamin，Scott Cuomo和Chris Whyte。 "与雅典娜作战：如何制造军队
Smarter, Faster, and More Efficient with Artificial Intelligence.” War on the Rocks. June 5, 2018. 人工智能更智能、更快、更高效。"岩石上的战争。 2018年6月5日。 https://warontherock
https://warontherocks.com/2018/06/wargaming-with-athena-how-to-make-militaries- s.com/2018/06/wargaming-with-athena-how-to-make-militaries 更智能、更快、更高效
smarter-faster-and-more-efficient-with-artificial-intelligence/. 的人工智能/。
Johnson, Bonnie, and William Treadway. “Artificial Intelligence—An Enabler of Naval Tactical Decision 约翰逊，邦妮和威廉*特德威。 "人工智能-海军战术决策的推动者
Superiority.” AI Magazine 40, no. 1 (Mar. 2019). 优越感。"AI杂志40,no.1(Mar. 2019).
Kabudi, Tumaini, Ilias Pappas, and Dag Olsen. “AI-Enabled Adaptive Learning Systems: A Systematic Kabudi，Tumaini，Ilias Pappas和Dag Olsen。 "人工智能自适应学习系统：系统化
Mapping of the Literature.” Computers and Education: Artificial Intelligence 2 (2021). 文献的映射。"计算机与教育：人工智能2（2021）。
Konaev, Margarita and Samuel Bendett. “Russian AI-Enabled Combat: Coming to a City near You?” War 科纳耶夫玛格丽塔和塞缪尔*本德特 "俄罗斯AI启用战斗：来到你附近的城市？"岩石上的战争。 201
on the Rocks. July 31, 2019. https://warontherocks.com/2019/07/russian-ai-enabled-combat- 9年7月31日。 https://warontherocks.com/2019/07/russian-ai-enabled-combatcoming-to-a-city
coming-to-a-city-near-you/. -near-you/。
Langville, Amy, and Carl Meyer. Google's PageRank and Beyond. Princeton, NJ: Princeton University 兰维尔，艾米和卡尔迈耶。谷歌的PageRank和超越。新泽西州普林斯顿：普林斯顿大学
Press, 2006. Press, 2006.
Laskey, Kathryn B. “Crowdsourced Decision Support for Emergency Responders.” In 18th International Laskey，Kathryn B."紧急响应者的众包决策支持。"在第18届国际
Command and Control Research and Technology Symposium, Alexandria, VA, June 19-21, 2013. 指挥和控制研究与技术研讨会，弗吉尼亚州亚历山大，2013年6月19-21日。 https://apps.d
https://apps.dtic.mil/sti/pdfs/ADA588344.pdf. tic.mil/sti/pdfs/ADA588344.pdf ...
Lehman, Joel, Jeff Clune, Dusan Misevic, Christoph Adami, Lee Altenberg, Julie Beaulieu, Peter Bentley, Lehman，Joel，Jeff Clune，Dusan Misevic，Christoph Adami，Lee Altenberg，Julie Beaulieu，P
et al. “The Surprising Creativity of Digital Evolution: A Collection of Anecdotes from the eter Bentley，et al。 "数字进化令人惊讶的创造力：来自进化计算和人造生命研究社区的轶事集合
Evolutionary Computation and Artificial Life Research Communities.” Nov. 21, 2019. 。"十一月。 21, 2019. arXiv：1803.03453。
arXiv:1803.03453.
65 65
Lewis, Larry, and Diane Vavrichek. Rethinking the Drone War. Quantico, VA: Marine Corps University 刘易斯，拉里和黛安Vavrichek。重新思考无人机战争。匡蒂科,VA:海军陆战队大学
Larry Lewis, AI Safety: An Action Plan for the Navy. CNA, DOP-2019-U-021957-1Rev, October 2019. 拉里*刘易斯，AI安全：海军行动计划。 CNA，DOP-2019-U-021957-1Rev，2019年10月。
Lewis, Larry. Drone Strikes: Civilian Casualty Considerations. Joint Staff. June 18, 2013. 刘易斯，拉里。无人机袭击：平民伤亡考虑。联合工作人员。 2013年6月18日。
Lewis, Larry. Insights for the Third Offset. CNA. DRM-2017-U-016281-Final. Sept. 2017. 刘易斯，拉里。第三偏移的见解。 CNA。 DRM-2017-U-016281-最终. "九月。 2017.
Lewis, Larry. Learning from Yemen. CNA. IRM-2019-U-019749-Final. May 2019. 刘易斯，拉里。向也门学习。 CNA。 IRM-2019-U-019749-决赛. 2019年5月。
https://www.cna.org/CNA_files/PDF/IRM-2019-U-019749-Final.pdf. https://www.cna.org/CNA_files/PDF/IRM-2019-U-019749-Final.pdf.
Lewis, Larry. Protecting Medical Care in Conflict: A Solvable Problem. CNA. Oct. 30, 2019. 刘易斯，拉里。在冲突中保护医疗：一个可以解决的问题。 CNA。十月。 30, 2019.
https://humanrightscommission.house.gov/sites/humanrightscommission.house.gov/files/doc https://humanrightscommission.house.gov/sites/humanrightscommission.house.gov/files/doc
uments/Protecting%20Medical%20Care%20in%20Conflict%20-%20Lewis.pdf. uments/Protecting%20Medical%20Care%20in%20Conflict%20-%20Lewis.pdf.
Lewis, Larry. Reducing and Mitigating Civilian Casualties: Enduring Lessons. Joint Staff. Apr. 12, 2013. 刘易斯，拉里。减少和减轻平民伤亡：持久的教训。联合工作人员。四月。 12, 2013.
Li, Manling, Alireza Zareian, Ying Lin, Xiaoman Pan, Spencer Whitehead, Brian Chen, Bo Wu, et al. Li,Manling,Alireza Zareian,Ying Lin,Xiaoman Pan,Spencer Whitehead,Brian Chen,Bo Wu,et al.
“GAIA: A Fine-Grained Multimedia Knowledge Extraction System.” In Proceedings of the 58th "GAIA：细粒度的多媒体知识提取系统。"In Proceedings Of The58th Annual Meeting Of The
Annual Meeting of the Association for Computational Linguistics: System Demonstrations, July Association For Computational Linguistics:System Demonstrations,July2020. https://aclant
2020. https://aclanthology.org/2020.acl-demos.11v2.pdf. hology.org/2020.acl-demos.11v2.pdf ...
Lopez, C. Todd. “Where It Counts, U.S. Leads in Artificial Intelligence.” DOD website. July 9, 2020. 洛佩兹，C*托德。 "重要的是，美国在人工智能方面处于领先地位。"国防部网站。 2020年7月9日。
https://www.defense.gov/Explore/News/Article/Article/2269200/where-it-counts-us-leads- https://www.defense.gov/Explore/News/Article/Article/2269200/where-it-counts-us-leads-
in-artificial-intelligence/. in-artificial-intelligence/.
Lowman, Christopher, and Gerald Prater. “Expansion of the Reconnaissance and Security BCT into the Lowman，Christopher和Gerald Prater。 "侦察和安全BCT扩展到
Cyber Domain: Lessons Learned from NTC Rotation 17-07.05.” July 2017. Unpublished white 网络领域：从NTC轮换中吸取的经验教训17-07.05。"2017年7月。未发表的白皮书。
paper.
Ludwig, Jeremy, and Bart Presnell. “Developing an Adaptive Opponent for Tactical Training.” In First 路德维希，杰里米和巴特普雷斯内尔。 "开发战术训练的适应性对手。"在第一
International Conference on Adaptive Instructional Systems (AIS), Orlando, FL, July 26–31, 适应性教学系统国际会议（AIS），奥兰多，佛罗里达州，2019年7月26-31日。
2019.
Marquis, Pierre, Odile Papini, and Henri Prade, eds. A Guided Tour of Artificial Intelligence Research: Marquis，Pierre，Odile Papini和Henri Prade，eds。人工智能研究导赏团:
Volume I: Knowledge Representation, Reasoning and Learning. Springer-Verlag, May 2020. 第一卷：知识表示、推理与学习。 Springer-Verlag，2020年5月。
McGrady, ED, and Justin Peachy, eds. Representing Artificial Intelligence in Wargames, Connections McGrady，ED和Justin Peachy，eds。在游戏、连接中代表人工智能
Conference Working Group 2 Proceedings, Dec. 2020. 会议第二工作组会议记录,12月12日。 2020. https://paxsims.files.wordpress.com/
https://paxsims.files.wordpress.com/2020/12/connections_wg2_2020_final-4a.pdf. 2020/12/connections_wg2_2020_final-4a.pdf ...
Meron, Theodor. “The Martens Clause, Principles of Humanity, and Dictates of Public Conscience.” The 梅隆，西奥多。 "马滕斯条款，人道原则和公众良知的命令。"的
American Journal of International Law 94, no. 1 (Jan. 2000). 美国国际法杂志94,第1期(1月1日)。 2000).
Metzler, Christopher A., David B. Lindell, and Gordon Wetzstein. “Keyhole Imaging: Non-Line-of-Sight Metzler，Christopher A.，David B.Lindell和Gordon Wetzstein。 "钥匙孔成像：非视线
Imaging and Tracking of Moving Objects Along a Single Optical Path.” IEEE Transactions on 沿单个光路对运动物体进行成像和跟踪。"IEEE交易
Computational Imaging 7 (Dec. 22, 2020). 计算成像7（Dec。 22, 2020).
Miatliuk, Kanstantsin, Adam Wolniakowski, and Pawel Kolosowski. “Engineering System of Systems Miatliuk，Kanstantsin，Adam Wolniakowski和Pawel Kolosowski。 "系统的工程系统
Conceptual Design in Theoretical Basis of Hierarchical Systems.” In 2020 IEEE International 分层系统理论基础的概念设计."在2020年IEEE系统、人和控制论国际会议（SMC），多伦
Conference on Systems, Man, and Cybernetics (SMC), Toronto, Ontario, Canada, Oct. 11–14, 多，安大略省，加拿大，10月。 11-14,2020,IEEE.
2020, IEEE.
66 66
Michalewicz, Zbigniew, and David B. Fogel. How to Solve It: Modern Heuristics. New York: Springer- Michalewicz，Zbigniew和David B.Fogel。如何解决它：现代启发式。纽约：斯普林格
Verlag, 2005. Verlag, 2005.
Michel, Arthur Holland. Unarmed and Dangerous: The Lethal Applications of Non-Weaponized Drones. 米歇尔，亚瑟*霍兰德。非武装和危险：非武器化无人机的致命应用。
The Center for the Study of the Drone at Bard College. Mar. 2020. 巴德学院无人机研究中心。玛尔。 2020. https://dronecenter.bard.edu/files/2020/
https://dronecenter.bard.edu/files/2020/03/CSD-Unarmed-and-Dangerous-Web.pdf. 03/CSD-Unarmed-and-Dangerous-Web.pdf ...
Mobus, George, and Michael Kalton. Principles of Systems Science (Understanding Complex Systems). 莫布斯，乔治和迈克尔*卡尔顿。系统科学原理（理解复杂系统）。
Springer, 2016. Springer, 2016.
Molinara, Mario, Alessandro Bria, Saverio De Vito, and Claudio Marrocco, eds. “Artificial Intelligence for Molinara，Mario，Alessandro Bria，Saverio De Vito和Claudio Marrocco，eds。 "人工智能为
Distributed Smart Sensing.” Special issue, Pattern Recognition Letters (Jan. 2021). 分布式智能传感。"特刊，模式识别信件（1月1日）。 2021).
Montoya, Jose, and Ricard Sole, “Small World Patterns in Food Webs.” Journal of Theoretical Biology Montoya，Jose和Ricard Sole，"食物网中的小世界模式。"理论生物学杂志214（2002）。
214 (2002).
Montoya, Jose, Stuart Pimm, and Ricard Sole. “Ecological Networks and Their Fragility.” Nature 442 Montoya，Jose，Stuart Pimm和Ricard Sole。 "生态网络及其脆弱性。"自然442（2006）。
(2006).
Moore, Andrew, Martial Hebert, and Shane Shaneman. “The AI Stack: A Blueprint for Developing and Moore，Andrew，Martial Hebert和Shane Shaneman。 "人工智能堆栈：发展和发展的蓝图
Deploying Artificial Intelligence.” In Ground/Air Multisensor Interoperability, Integration, and 部署人工智能。"在地面/空中多传感器互操作性、集成和
Networking for Persistent ISR IX, Proceedings Vol. 10635, May 4, 2018. 持续ISR IX的网络，会议记录卷。 10635，2018年5月4日。
Mourabit, Yousef, Youssef El Habouz, Mustapha Lydiri, and Hicham Zougagh. “A New Sentiment Mourabit，Yousef，Youssef El Habouz，Mustapha Lydiri和Hicham Zougagh。 "一种新的情绪
Analysis System of Tweets Based on Machine Learning Approach.” International Journal of 基于机器学习方法的推文分析系统。"国际期刊
Scientific and Technology Research 9, no. 12 (Dec. 2020). 科学与技术研究9,no.12(Dec. 2020).
Moy, Glenn, and Slava Shekh. “The Application of AlphaZero to Wargaming.” In Advances in Artificial Moy，Glenn和Slava Shekh。 "AlphaZero对Wargaming的应用。"在人造的进步
Intelligence, Proceedings of 32nd Australasian Joint Conference, Adelaide, SA, Australia, Dec. 情报，第32届澳大拉西亚联席会议记录，阿德莱德，澳大利亚，12月。 2019.
2019.
Muggleton, Stephen H., and Hiroaki Watanabe, eds. Latest Advances in Inductive Logic Programming. Muggleton，Stephen H.和Hiroaki Watanabe，eds。电感逻辑编程的最新进展.
London: Imperial College Press, 2014. 伦敦：帝国理工学院出版社，2014年。
NATO/ISAF, Tactical Directive, July 6, 2009. 北约/安援部队，战术指令，2009年7月6日。

https://www.nato.int/isaf/docu/official_texts/Tactical_Directive_090706.pdf. https://www.nato.int/isaf/docu/official_texts/Tactical_Directive_090706.pdf.
NSCAI. Final Report. Mar. 2021. https://www.nscai.gov/wp-content/uploads/2021/03/Full-Report- NSCAI。最后报告。玛尔。 2021. https://www.nscai.gov/wp-content/uploads/2021/03/Full-Report-
Digital-1.pdf. Digital-1.pdf.
“OFFensive Swarm-Enabled Tactics (OFFSET).” DARPA website. "进攻群启用战术（偏移）。"DARPA网站。
https://www.darpa.mil/program/offensive-swarm-enabled-tactics. https://www.darpa.mil/program/offensive-swarm-enabled-tactics.
Paltrinieri, Nicola, Louise Comfort, and Genserik Reniers. “Learning About Risk: Machine Learning for Paltrinieri，Nicola，Louise Comfort和Genserik Reniers。 "了解风险：机器学习
Risk Assessment.” Safety Science 118 (Oct. 2019). 风险评估。"安全科学118（Oct。 2019).
Pearl, Judea. “The Seven Tools of Causal Inference, with Reflections on Machine Learning.” 珀尔，朱迪亚。 "因果推理的七种工具，对机器学习的反思。”
Communications of the ACM 62, no. 3 (Mar. 2019). ACM62,第3号通讯(3月. 2019).
Pellerin, Cheryl. “Project Maven to Deploy Computer Algorithms to War Zone by Year’s End.” DOD 佩勒林，谢丽尔。 "项目Maven到年底将计算机算法部署到战区。"国防部网站。 2017年7月21日
website. July 21, 2017. 。 https://www.defense.gov/Explore/News/Article/Article/1254719/project-maven-to-deployc
https://www.defense.gov/Explore/News/Article/Article/1254719/project-maven-to-deploy- omputer-algorithms-to-war-zone-by-years-end/。
computer-algorithms-to-war-zone-by-years-end/.
67 67
Pomerleau, Mark. “Air Force Software Tool Helped Coordinate Afghanistan Evacuation of Civilians.” Pomerleau，马克。 "空军软件工具帮助协调阿富汗平民撤离。”
C4ISRNet. Sept. 23, 2021. https://www.c4isrnet.com/battlefield-tech/c2- C4ISRNet. Sept. 23, 2021. https://www.c4isrnet.com/battlefield-tech/c2-
comms/2021/09/23/air-force-software-tool-helped-coordinate-afghanistan-evacuation-of- comms/2021/09/23/air-force-software-tool-helped-coordinate-afghanistan-evacuation-of-
civilians/. civilians/.
Popescu, Dan, Florian Stoican, Grigore Stamatescu, Oana Chenaru, and Loretta Ichim. “A Survey of Popescu，Dan，Florian Stoican，Grigore Stamatescu，Oana Chenaru和Loretta Ichim。 "调查
Collaborative UAV–WSN Systems for Efficient Monitoring.” Sensors 19, no. 21 (Oct. 28, 2019). 协作无人机-WSN系统进行高效监控."传感器19号，第21号（10月19日）。 28, 2019). ht
https://www.mdpi.com/1424-8220/19/21/4690/htm. tps://www.mdpi.com/1424-8220/19/21/4690/htm ...
Ranaei, Samira, and Arho Suominen. “Using Machine Learning Approaches to Identify Emergence: Case Ranaei，Samira和Arho Suominen。 "使用机器学习方法来识别出现：车辆相关专利数据的案例。"2
of Vehicle Related Patent Data.” In 2017 Portland International Conference on Management of 017年波特兰工程与技术管理国际会议（PICMET），波特兰，OR，2017年7月9-13日，IEEE。
Engineering and Technology (PICMET), Portland, OR, July 9-13, 2017, IEEE.
“Real-Time Airspace Awareness and De-Confliction for Future Battles.” DARPA website. Apr. 7, 2020. "未来战斗的实时空域意识和消除冲突。"DARPA网站。四月。 7, 2020.
https://www.darpa.mil/news-events/2020-04-07. https://www.darpa.mil/news-events/2020-04-07.
Regens, James L. “Augmenting Human Cognition to Enhance Strategic, Operational, and Tactical Regens，James L."增强人类认知以增强战略，操作和战术
Intelligence.” Intelligence and National Security 34, no. 5 (2019). 智力。"情报与国家安全34，no.5（2019）。
Reinaldo, Francisco, Marcus Siqueira, Rui Camacho, and Luis Paulo Reis. “A Tool for Multi-Strategy Reinaldo，Francisco，Marcus Siqueira，Rui Camacho和Luis Paulo Reis。 "多策略的工具
Learning.” Advances in AI Research in Computing Science 26 (2006). 学习。"Advances in AI Research in Computing Science26（2006）。
“Roborder: Aims & Objectives.” Roborder website. https://roborder.eu/the-project/aims-objectives/. "Roborder：目标和目标。"Roborder网站。 https://roborder.eu/the-project/aims-objectives/。

“Roborder: Autonomous Swarm of Heterogeneous Robots for Border Surveillance.” Roborder website. "Roborder：用于边境监视的异构机器人的自主群。"Roborder网站。访问9月。 20, 2021. https://r
Accessed Sept. 20, 2021. https://roborder.eu/wp- oborder.eu/wpcontent/uploads/2021/03/ROBORDER_General_v2.0.pdf ...
content/uploads/2021/03/ROBORDER_General_v2.0.pdf.
Russell, Stuart, and Peter Norvig. Artificial Intelligence: A Modern Approach, 4th Edition. Pearson, May 拉塞尔，斯图尔特和彼得诺维格。人工智能：现代方法，第4版。皮尔逊，2021年5月。
2021.
Rutherford, Alex, Manuel Cebrian, Inho Hong, and Iyad Rahwan. “Impossible by Conventional Means: 卢瑟福，亚历克斯，Manuel Cebrian，Inho Hong和Iyad Rahwan。 "通过常规手段是不可能的:
Ten Years on from the DARPA Red Balloon Challenge.” Aug. 13, 2020. arXiv:2008.05940v1. 十年后，DARPA红气球挑战赛。"八月。 13, 2020. arXiv：2008.05940v1。
Saputra, Muhammad, Andrew Markham, and Niki Trigoni. “Visual SLAM (Simultaneous Localization Saputra，Muhammad，Andrew Markham和Niki Trigoni。 "视觉SLAM（同时本地化
and Mapping) and Structure from Motion in Dynamic Environments: A Survey.” ACM Computing and Mapping)and Structure from Motion in Dynamic Environments:A Survey."ACM计算调查5
Surveys 51, no. 2 (June 2018). 1，第2号（2018年6月）。
Saunders, Charles, John Murray-Bruce, and Vivek Goyal. “Computational Periscopy with an Ordinary 桑德斯，查尔斯，约翰默里-布鲁斯和维维克戈亚尔。 "用普通的计算潜望镜
Digital Camera.” Nature 365 (Jan. 23, 2019). https://www.nature.com/articles/s41586-018- 数码相机。"自然365（1月。 23, 2019). https://www.nature.com/articles/s41586-018 0
0868-6. 868-6．
Schedl, David C., Indrajit Kurmi, and Oliver Bimber. “Search and Rescue with Airborne Optical Schedl，David C.，Indrajit Kurmi和Oliver Bimber。 "用机载光学搜索和救援
Sectioning.” Nature Machine Intelligence 2 (Nov. 23, 2020). 分割。"自然机器智能2（Nov。 23, 2020).
Schrittwieser, Julian, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schrittwieser,Julian,Ioannis Antonoglou,Thomas Hubert,Karen Simonyan,Laurent Sifre,Simon
Schmitt, Arthur Guez, et al. “Mastering Atari, Go, Chess and Shogi by Planning with a Learned Schmitt，Arthur Guez，et al。通过有学问的计划，掌握Atari、围棋、国际象棋和幕府
Model.” Nov. 19, 2019. https://arxiv.org/abs/1911.08265. Model.” Nov. 19, 2019. https://arxiv.org/abs/1911.08265.
Schwartz, Peter J., Daniel O’Neill, Meghan Bentz, Adam Brown, Brian Doyle, Olivia Liepa, Robert Schwartz，Peter J.，Daniel O'Neill，Meghan Bentz，Adam Brown，Brian Doyle，Olivia Liepa，Robert
Lawrence, et al. “AI-Enabled Wargaming in the Military Decision Making Process.” In Artificial 劳伦斯等人。 "AI在军事决策过程中启用了Wargaming。"在人工智能和机器学习的多领域
Intelligence and Machine Learning for Multi-Domain Operations Applications II, Proceedings 操作应用II，会议记录
Vol. 11413, 2020. Vol. 11413, 2020.
68 68
Shrestha, Yash, Shiko Ben-Mehahem, and Georg von Krogh. “Organizational Decision-Making Shrestha，Yash，Shiko Ben-Mehahem和Georg von Krogh。 "组织决策
Structures in the Age of Artificial Intelligence.” California Management Review 61, no. 4 (Aug. 人工智能时代的结构。"加州管理评论61，第4期（8月）。 2019).
2019).
Silver, David, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Silver,David,Thomas Hubert,Julian Schrittwieser,Ioannis Antonoglou,Matthew Lai,Arthur Guez,
Marc Lanctot, et al. “Mastering Chess and Shogi by Self-Play with a General Reinforcement Marc Lanctot等人。 "掌握国际象棋和将棋的自我发挥与一般强化
Learning Algorithm.” Dec. 5, 2017. arXiv:1712.01815v1. 学习算法。"十二月。 5, 2017. arXiv：1712.01815v1。
Soldin, Ryan J. “SAR Target Recognition with Deep Learning.” In IEEE Applied Imagery Pattern Soldin，Ryan J."Sar目标识别与深度学习。"在IEEE应用图像模式
Recognition Workshop (AIPR), Washington, DC, Oct. 2018. 认可工作坊(AIPR),华盛顿特区,10月10日. 2018.
Sottilare, Robert. “A Hybrid Machine Learning Approach to Automated Scenario Generation (ASG) to 索蒂拉雷，罗伯特。 "自动化场景生成（ASG）的混合机器学习方法
Support Adaptive Instruction in Virtual Simulations and Games.” In I3M Defense & Homeland 支持虚拟模拟和游戏中的自适应指令."在I3M国防与家园
Security Simulation Workshop, Budapest, Hungary, 2018. 安全模拟研讨会，匈牙利布达佩斯，2018。
Stewart, Phil, and Idrees Ali. “U.S. Says Kabul Drone Strike Killed 10 Civilians, Including Children, in 斯图尔特，菲尔和伊德雷斯阿里。 "美国说，喀布尔无人机袭击造成包括儿童在内的10平民在'悲
‘Tragic Mistake.’” Reuters. Sept. 18, 2021. https://www.reuters.com/world/asia-pacific/us- 惨的错误中丧生。'"路透社。 "九月。 18, 2021. https://www.reuters.com/world/asia-pacific/us
military-says-10-civilians-killed-kabul-drone-strike-last-month-2021-09-17/. military-says-10-civilians-killed-kabul-drone-strike-last-month-2021-09-17/。
Strum, Luke. “Intel Airmen Sharpen AI Technology for Domestic Response.” Air National Guard 斯特鲁姆，卢克。 "英特尔飞行员锐化AI技术用于国内反应。"空军国民警卫队
website. Nov. 25, 2019. https://www.ang.af.mil/Media/Article-Display/Article/2025332/intel- 网站。 11月11日。 25, 2019. https://www.ang.af.mil/Media/Article-Display/Article/20253
airmen-sharpen-ai-%20technology-for-domestic-response/. 32/intel airmen-sharpen-ai-%20technology-for-domestic-response/.
Sutton, Richard S., and Andrew G. Barto. Reinforcement Learning: An Introduction. Cambridge, MA: MIT 萨顿，理查德S.和安德鲁G.巴托。强化学习:介绍. 马萨诸塞州剑桥：麻省理工学院
Tadjdeh, Yasmin. “China Threatens U.S. Primacy in Artificial Intelligence (UPDATED).” National Tadjdeh，Yasmin。 "中国威胁美国在人工智能方面的首要地位（更新）。"国家
Defense. Oct. 31, 2020. https://www.nationaldefensemagazine.org/articles/2020/10/30/china- 辩护。十月。 31, 2020. https://www.nationaldefensemagazine.org/articles/2020/10/30/ch
threatens-us-primacy-in-artificial-intelligence. ina 威胁到美国在人工智能中的首要地位。
Tan, Qiaoyu, Ninghao Liu, and Xia Hu. “Deep Representation Learning for Social Network Analysis.” 谭，乔宇，刘宁浩，夏胡。 "社交网络分析的深度表示学习。”
Apr. 18, 2019. arXiv:1904.08547 [cs.SI]. Apr. 18, 2019. arXiv:1904.08547 [cs.SI].
Tang, John C., Manuel Cebrian, Nicklaus A. Giacobe, Hyun-Woo Kim, Taemie Kim, and Douglas Wickert. Tang，John C.，Manuel Cebrian，Nicklaus A.Giacobe，Hyun-Woo Kim，Taemie Kim和Douglas Wi
“Reflecting on the DARPA Red Balloon Challenge.” Communications of the ACM 54, no. 4 (Apr. ckert。 "反思DARPA红气球挑战赛。"ACM54通讯，第4号（4月）。 2011). https://cacm.acm.org/
2011). https://cacm.acm.org/magazines/2011/4/106587-reflecting-on-the-darpa-red-balloon- magazines/2011/4/106587-reflecting-on-the-darpa-red-balloonchallenge/fulltext ...
challenge/fulltext.
Taylor, Curt. “It’s Time for Cavalry to Get Serious about Cyber Reconnaissance.” Armor: Mounted 泰勒，柯特。 "现在是骑兵认真对待网络侦察的时候了。"装甲：装
Maneuver Journal (Fall 2018). 机动杂志（2018年秋季）。 https://mcoe.azurewebsites.us/Armor/eARMOR/content/iss
https://mcoe.azurewebsites.us/Armor/eARMOR/content/issues/2018/Fall/4Taylor18.pdf. ues/2018/Fall/4Taylor18.pdf ...
The White House. July 1, 2016. Exective Order--United States Policy on Pre- and Post-Strike Measures to 白宫。 2016年7月1日。 Exective Order--美国关于罢工前和罢工后措施的政策
Address Civilian Casualties in U.S. Operations Involving the Use of Force. 处理涉及使用武力的美国行动中的平民伤亡。
Torfi, Amirsina, Rouzbeh Shirvani, Yaser Keneshloo, Nader Tavaf, and Edward Fox. “Natural Language Torfi，Amirsina，Rouzbeh Shirvani，Yaser Keneshloo，Nader Tavaf和Edward Fox。 "自然语言
Processing Advancements by Deep Learning: A Survey.” Feb. 27, 2021. arXiv:2003.01200v4 通过深入学习处理进步:一项调查."二月。 27, 2021. arXiv：2003.01200v4[cs.CL]。
[cs.CL].
Torreno, Alejandro, Eva Onaindia, Antonin Komenda, and Michal Stolba. “Cooperative Multi-Agent Torreno，Alejandro，Eva Onaindia，Antonin Komenda和Michal Stolba。 "合作多代理
Planning: A Survey.” ACM Computing Surveys 50, no. 6 (Nov. 2017). 规划：调查。"ACM计算调查50号，第6号（11月11日）。 2017).
69 69
Toth, Susan, and William Hughes. “The Journey to Collaborative AI at the Tactical Edge (CATE).” In 托特，苏珊和威廉*休斯。 "在战术边缘（CATE）的协作人工智能之旅。"在
Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications III, 人工智能和机器学习的多领域操作应用III,
Proceedings Vol. 11746, Apr. 12, 2021. 诉讼程序卷。 11746年4月 12, 2021.
Trehan, Robin. “How AI Is Transforming Fraud Prevention in Banking and Finance.” Deltec. May 11, 特雷汉，罗宾。 "人工智能如何改变银行和金融领域的欺诈预防。"德尔特克。 2020年5月11日。 h
2020. https://www.deltecbank.com/2020/05/11/how-ai-is-transforming-risk-in-finance-and- ttps://www.deltecbank.com/2020/05/11/how-ai-is-transforming-risk-in-finance-andbanking/?l
banking/?locale=en. ocale=en ...
Trent, Stoney, and Scott Lathrop. “A Primer on Artificial Intelligence for Military Leaders.” Small Wars 特伦特，斯托尼和斯科特*莱瑟罗普。 "军事领导人人工智能入门。"小型战争
Journal (Aug. 2018). Journal (Aug. 2018).
UK Ministry of Defence. SAPIENT Middleware Interface Control Document. May 11, 2020. 英国国防部。 SAPIENT中间件接口控制文档。 2020年5月11日。
https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_da https://assets.publishing.service.gov.uk/government/uploads/system/uploads/attachment_da
ta/file/895246/SAPIENT_Interface_Control_Document_v5.0.pdf. ta/file/895246/SAPIENT_Interface_Control_Document_v5.0.pdf.
Ushahidi homepage. https://www.ushahidi.com/. Ushahidi homepage. https://www.ushahidi.com/.
US Air Force. C2IMERA Overview. Kessel Run. Mar. 26, 2020. Video. 美国空军。 C2IMERA概述。凯塞尔跑。玛尔。 26, 2020. 视频。
https://www.youtube.com/watch?v=8meuTt1RQbE. https://www.youtube.com/watch?v=8meuTt1RQbE.
US Air Force. Kessel Run homepage. https://kesselrun.af.mil/. 美国空军。凯塞尔运行主页。 https://kesselrun.af.mil/。
Vatsa, Mayank, Richal Singh, and Angshul Majumdar, eds. Deep Learning in Biometrics. Boca Raton, FL: Vatsa，Mayank，Richal Singh和Angshul Majumdar，eds。生物测定学的深入学习。博卡拉顿,FL:
CRC Press, 2018. CRC Press, 2018.
Vinyals, Oriol, Igor Babuschkin, Wojciech Czarnecki, Michael Mathieu, Andrew Didzik, Junyoung Chung, Vinyals,Oriol,Igor Babuschkin,Wojciech Czarnecki,Michael Mathieu,Andrew Didzik,Junyoung Chung,
David Choi, et al. “Grandmaster Level in StarCraft II Using Multi-Agent Reinforcement Learning.” David Choi等人。 "星际争霸II中的特级大师级使用多代理强化学习。”
Nature 575 (Oct. 30, 2019). https://www.nature.com/articles/s41586-019-1724-z. Nature 575 (Oct. 30, 2019). https://www.nature.com/articles/s41586-019-1724-z.
Wilensky, Uri, and William Rand. An Introduction to Agent-Based Modeling. Cambridge, MA: MIT Press, 威伦斯基，乌里和威廉兰德。基于代理的建模简介。雁抚蜷医帑九＆缧
2015.
Work, Robert. Memorandum. Subject: Establishment of an Algorithmic Warfare Cross-Functional Team 工作，罗伯特。备忘录。主题：建立算法战跨职能团队（Project Maven）。四月。 26, 2017. http
(Project Maven). Apr. 26, 2017. s://dodcio.defense.gov/Portals/0/Documents/Project%20Maven%20DSD%20Memo%202 0170425
https://dodcio.defense.gov/Portals/0/Documents/Project%20Maven%20DSD%20Memo%202 ．pdf。
0170425.pdf.
Wray, Robert E., James Kirk, and John Laird. “Language Models as a Knowledge Source for Cognitive Wray，Robert E.，James Kirk和John Laird。 "语言模型作为认知的知识来源
Agents.” Sept. 20, 2021. arXiv:2109.08270v2 [cs.AI]. Agents.” Sept. 20, 2021. arXiv:2109.08270v2 [cs.AI].
Xiang, Tian-Zhu, Gui-Song Xia, and Liangpei Zhang. “Mini-Unmanned Aerial Vehicle-Based Remote 向、天珠、桂松夏和张良培。 "基于微型无人飞行器的远程
Sensing: Techniques, Applications, and Prospects.” IEEE Geoscience and Remote Sensing 传感：技术、应用和前景。"IEEE地球科学与遥感
Magazine 7, no. 3 (Sept. 2019). 杂志7，第3期（9月）。 2019).
Yusuf, Sagir, and Chris Baber. “Inferential Reasoning for Heterogeneous Multi-Agent Mission.” 优素福，萨吉尔和克里斯巴伯。 "异构多代理任务的推理。”
International Scholarly and Scientific Research & Innovation 14, no. 10 (2020). 国际学术与科学研究与创新14,第10期(2020).
Yusuf, Salisu Ibrahim, Steve Adeshina, and Moussa Mahamat Boukar. “Parameters for Human Gait 优素福，Salisu Ibrahim，Steve Adeshina和Moussa Mahamat Boukar。 "人体步态参数
Analysis: A Review.” In 2019 15th International Conference on Electronics, Computer and 分析：回顾。"2019年第十五届电子、计算机和国际会议"
Computation (ICECCO), Dec. 2019. 计算(ICECCO),12月. 2019.
70 70
Zhao, Mingmin, Tianhong Li, Mohammad Abu Alsheikh, Yonglong Tian, Hang Zhao, Antonio Torralba, 赵、明敏、李天宏、穆罕默德*阿布*阿尔谢克、永隆*田、杭*赵、安东尼奥*托拉尔巴和迪娜*卡塔
and Dina Katabi. “Through-Wall Human Pose Estimation Using Radio Signals.” Computer Vision 比。 "使用无线电信号进行穿墙人体姿势估计。"计算机视觉和模式识别（CVPR）（2018）。 http:
and Pattern Recognition (CVPR) (2018). //openaccess.thecvf.com/content_cvpr_2018/papers/Zhao_ThroughWall_Human_Pose_CVPR_20
http://openaccess.thecvf.com/content_cvpr_2018/papers/Zhao_Through- 18_paper.pdf ...
Wall_Human_Pose_CVPR_2018_paper.pdf.
Zhen, Yang, Zhang Wanpeng, and Liu Hongfu. “Real-Time Strategy Game Tactical Recommendation 甄、杨、张万鹏、刘洪福。 "即时战略游戏战术推荐
Based on Bayesian Network.” Journal of Physics: Conference Series 1168, no. 3 (2019). 基于贝叶斯网络。"物理学杂志：会议系列1168，第3期（2019）。
Zhu, Xiaoyu, Junwei Liang, and Alexander Hauptmann. “MSNet: A Multilevel Instance Segmentation 朱，小雨，Junwei Liang和Alexander Hauptmann。 "MSNet：多级实例分段
Network for Natural Disaster Damage Assessment in Aerial Videos.” Dec. 30, 2020. 航拍视频自然灾害损害评估网络."十二月。 30, 2020. arXiv：2006.16479[cs.CV
arXiv:2006.16479 [cs.CV]. ]。
71 71
This report was written by CNA’s Operational Warfighting Division 这份报告是由中央社作战部撰写的
(OPS). (OPS).
OPS focuses on ensuring that US military forces are able to compete and OPS专注于确保美国军队能够竞争和
win against the nation’s most capable adversaries. The major functional 战胜国家最有能力的对手. 主要功能
components of OPS work include activities associated with generating 项目事务处工作的组成部分包括与生成相关的活动
and then employing the force. Force generation addresses how forces and 然后使用原力。力的产生解决了力和力的产生问题。
commands are organized, trained, scheduled, and deployed. Force 命令被组织、训练、计划和部署。警队
employment encompasses concepts for how capabilities are arrayed, 就业包括如何排列能力的概念,
protected, and sustained at the operational level in peacetime and 在和平时期的行动层面受到保护和维持。
conflict, in all domains, against different types of adversaries, and under 冲突,在所有领域,针对不同类型的对手,和下
varied geographic and environmental conditions. 不同的地理和环境条件。
CNA is a not-for-profit research organization that serves the public CNA是一个非营利性研究组织，通过提供深入的分析和以结果为导向的

interest by providing in-depth analysis and result-oriented solutions to 解决方案，帮助政府领导人选择制定政策和管理运营的最佳行动方案，
help government leaders choose the best course of action in setting 为公众利益
policy and managing operations.
DOP-2021-U-030953-1Rev-2 DOP-2021-U-030953-1Rev-2
3003 Washington Boulevard, Arlington, VA 22201 3003华盛顿大道，阿灵顿，VA22201

www.cna.org ● 703-824-2000 www.cna.org ● 703-824-2000

AD1181578 Merged Public

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AD1181578 Merged Public

Uploaded by

Copyright:

Available Formats

February 2022 February 2022

Leveraging AI to Mitigate Civilian Harm 利用人工智能减轻平民伤害

DISTRIBUTION STATEMENT A. Approved for public release: distribution unlimited. 发行声明A.批准公开发行:发行无限.

Public Release. 公开发布。

Dr. David Broyles, Director 戴维*布洛伊博士，董事

Request additional copies of this document through inquiries@cna.org. 透过以下途径索取本文件的额外副本inquiries@cna.org...

Copyright © CNA 2022 版权附图CNA2022

5c. PROGRAM ELEMENT NUMBER 5C.程序元素编号0605154N

12. DISTRIBUTION / AVAILABILITY STATEMENT 12. 发行/可用性声明发行声明A.批准公开发行：发行无限。

13. SUPPLEMENTARY NOTES 13. 补充说明

14. ABSTRACT 14. 摘要在过去十年中，关于将人工智能引入战争的争论非常激烈。

15. SUBJECT TERMS 15. 主题术语人工智能，AI，LOAC，自治，平民伤亡

CNA Occasional Paper | i CNA临时文件/i

• Alerting the presence of transient civilians by using object identification to • 通过使用物体识别自动监测目标区域周围的其他个人，并在检测到他们时发出警报

CNA Occasional Paper | ii 中央社临时文件/ii

CNA Occasional Paper | iii 中央社临时文件/iii

CNA Occasional Paper | iv CNA临时文件/iv

Connecting the CPLC with AI/ML applications .................................................................................... 23 将CPLC与AI/ML应用程序连接。................................................................................... 23

CNA Occasional Paper | v CNA临时文件/v

CNA Occasional Paper | vi CNA临时文件/vi

• AI for target identification, including identifying US military aircraft on runways • 用于目标识别的AI，包括识别跑道上的美国军用飞机（报告准确率为92％）和海上

Concerns regarding civilian harm risks 对平民伤害风险的关注

13 Yasmin Tadjdeh, “China Threatens U.S. Primacy.” 13Yasmin Tadjdeh，"中国威胁美国的首要地位。”

Opportunity for using AI to mitigate civilian 利用人工智能减轻平民伤害的机会

CNA’s work on civilian harm 中央社关于平民伤害的工作

Developing a methodology for understanding civilian harm 制定了解平民伤害的方法

Applying the methodology: Iraq and Afghanistan 应用方法：伊拉克和阿富汗

Iraq: Analysis of escalation of force 伊拉克：对武力升级的分析

Afghanistan: Real-time support Afghanistan: Real-time support

Tracking civilian harmcivilian harm 追踪平民危害

Monitoring and responding to trends 监测和应对趋势

24NATO/ISAF, Tactical Directive, July 6, 2009. 24北约/安援部队，战术指令，2009年7月6日。 https://www.nato.int/isaf/

Root cause determinations: Why civilian harm 根本原因确定：为什么会发生平民伤害

Source: CNA. Source: CNA.

A civilian protection life cycle for mitigating 减少风险的平民保护生命周期

Figure 2. Civilian protection life cycle 图2。 平民保护生命周期

Source: CNA. Source: CNA.

Examples of practical steps for mitigating 减轻平民伤害的实际步骤的例子

Detecting transient civilian presence 检测瞬态平民存在

Recognizing protected symbols 识别受保护的符号

Improving collateral damage estimates 改善附带损害估计

Detecting misassociations Detecting misassociations

Connecting the CPLC with AI/ML applications 连接CPLC与AI/ML应用

Source: CNA. Source: CNA.

Autonomous UASs/UAVs Autonomous UASs/UAVs

44 “Roborder: Aims & Objectives,” Roborder website, https://roborder.eu/the-project/aims-objectives/. 44"Roborder:Aims&Objectives,"Roborder网站,https://roborder.eu/the-project/aims-objectives/。

Computer Vision 计算机视觉

Knowledge Representation and Reasoning 知识表示与推理

Natural Language Processing 自然语言处理

Planning and Scheduling 计划及时间表

Threat Monitoring 威胁监控

Commentary about the CPLC-AI matrix in Figure 3 关于图3中CPLC-AI矩阵的评论

Artificial Intelligence (AI) / Machine Learning (ML) Applications 人工智能(AI)/机器学习(ML)应用

Cycle Assess 2 2 2 6 5 4 0 2 2 3 3 31 Cycle Assess 2 2 2 6 5 4 0 2 2 3 3 31

Source: CNA. Source: CNA.

137 Li et al., “GAIA.” p. 1. 137Li等人。，"盖亚。"第1页。

Reducing but not eliminating risks to civilians 减少但不消除对平民的风险

140 US Air Force, Kessel Run homepage, https://kesselrun.af.mil/. 140美国空军,凯塞尔运行主页,https://kesselrun.af.mil/。

Source: CNA. Source: CNA.

Source: CNA. Source: CNA.

Figure 6 adds four more layers of ambiguity and uncertainty: 图6增加了四层模糊性和不确定性:

Source: CNA. Source: CNA.

CIVCAS Cause-Mitigations-AI-applications CIVCAS Cause-Mitigations-AI-applications

Figure 2. Civilian protection life cycle 图2。平民保护生命周期