标签： ECON6025

经济代写|博弈论代写Game Theory代考|MultiLayered Defense Network Structure

Posted on 2023年4月13日2023年4月13日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

博弈论是对理性主体之间战略互动的数学模型的研究。它在社会科学的所有领域，以及逻辑学、系统科学和计算机科学中都有应用。最初，它针对的是两人的零和博弈，其中每个参与者的收益或损失都与其他参与者的收益或损失完全平衡。在21世纪，博弈论适用于广泛的行为关系；它现在是人类、动物以及计算机的逻辑决策科学的一个总称。

statistics-lab™ 为您的留学生涯保驾护航在代写博弈论Game Theory方面已经树立了自己的口碑, 保证靠谱, 高质且原创的统计Statistics代写服务。我们的专家在代写博弈论Game Theory代写方面经验极为丰富，各种代写博弈论Game Theory相关的作业也就用不着说。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|MultiLayered Defense Network Structure

We simply deploy HINs, MINs, and LINs to core, middle, and outer layers, correspondingly. To be specific, each layer consists of the corresponding types of nodes as follows:

The core layer consists of nodes with high importance (i.e. HINs). Since nodes in this layer have high importance levels given, if they are compromised and/or failed, it may easily lead to a system failure, as high importance nodes are weighted more in estimating the BF condition, as shown in Eq. (20.2). The example nodes may include central database servers or authentication servers in the cloud. The importance range of HINs is in [8-10] as an integer and each HIN’s importance value is assigned based on this range. Due to the high importance of nodes in this layer, all nodes in the core layer will be targeted by attackers. Recall that HINs can be directly connected only with either MINs or HINs, not LINs. This network topology is designed to add more security protection toward HINs in the core layer.
The middle layer consists of nodes with medium importance (i.e. MINs). The example nodes may include edge servers which primarily connect the server in the core layer to sensors and IoT devices in the outer layer, such as routers and switches. The importance range of MINs is set in [4-7] as an integer and each MIN’s importance is assigned based on this range. MINs can be directly connected with either HINs or LINs as they mainly serve as a coordinator between the core and outer layers. The failure or compromise of MINs may hinder service availability upon a lack of MINs, in addition to possibly causing a system failure due to triggering the BF condition in Eq. (20.2).
The outer layer consists of nodes with low importance (i.e. LINs). The example nodes may include low-cost IoT devices and/or sensors. Nodes in this layer (i.e. LINs) will have their importance ranged in $[1,3]$ as an integer. Similar to MINs, the compromise or failure of LINs may trigger the BF condition. LINs can be directly connected with LINs or MINs, but not HINs, as the LINs in the outer layer is the first layer attackers can access to penetrate into a given network. A lack of LINs may lead to a shallow protection layer, which may lead to easy access for the attackers to get into the deeper layers.

To maintain this priority-based, multilayered network structure, even if nodes move around, they need to check which nodes to communicate with while not violating the multilayered network structure. Figure 20.1 shows an example SD-IoBT network structure described in this section.

经济代写|博弈论代写Game Theory代考|Selection of an Intrusion Response Policy

Formulation of the MDP: In Markov Decision Process (MDP), we assume complete state information available for the environment and the next state of the environment only depends on the current state and the chosen action.

This work considers an MDP as a mathematical framework for DRL, by a tuple $\langle\delta, \mathscr{A}, \mathscr{F}, \mathscr{R}, \gamma\rangle$ for a DRL agent, consisting of (1) a set of states $\delta$ with a distribution of starting states $\mathrm{p}\left(\mathbf{s}0\right)$ where $\mathbf{S}_t$ refers to a set of states at time $t ;(2)$ a set of actions, $\mathscr{A}=\left{\mathbf{a}_1, \ldots, \mathbf{a}_t, \ldots \mathbf{a}_T\right}$ where $\mathbf{a}_t$ is a set of actions available at time $t$; (3) a transition probability, $\mathscr{T}_i\left(\mathbf{s}_l, \mathbf{a}_t, \mathbf{s}{t+1}\right)$, from state $\mathbf{s}l$ to state $\mathbf{s}{t+1}$ via $\mathbf{a}t ;$ (4) an immediate reward function, $\mathscr{R}\left(\mathbf{s}_t, \mathbf{a}_l, \mathbf{s}{l+1}\right)$; (5) a reward discount factor (e.g. $\gamma$ ); and (6) policy $\pi$ mapping from states to a probability distribution of actions, $\pi: \delta \rightarrow(\mathscr{A}=\mathbf{a} \mid \delta)$. An RL agent aims to maximize its own total expected return $\mathscr{R}=\sum_{t=0}^T \gamma_t R_t$ where $\gamma$ is a discounting factor and $T$ is the time horizon. The MDP can generate an optimal policy $\pi^*$ that maximizes the expected reward from states (Arulkumaran et al. 2017a).

To select an optimal intrusion response action that can maximize both security and performance of a given network, we formulate the MDP by:

States $(\delta)$ : A state at time $t$ by taking an action, denoted by $S_l$, refers to a system state in terms of system security and performance (i.e. $\mathrm{SS}_l$ and $\mathrm{PS}_l$ ), which are detailed later in this section. following three actions can be taken: destruction, repair, or replacement. Each action is detailed in Section 20.4.2. The reward associated with each action taken is formulated later in this section. ticular action. The DRL agent will select one of three actions.
Discount factor $(\gamma)$ : We will consider a discount factor, $\gamma$, ranged in $(0,1]$ as a real number.
Objectives of the Multi-Layered Network Structure: At each layer, an agent sitting on a corresponding SDN controller will keep track of the number of nodes detected as compromised/failed (NDCFs) and the number of messages correctly delivered in the layer it governs. Based on the information informed by each agent to the DRL agent, the DRL agent will estimate the system state. For simplicity, we consider system security state and performance states to represent a system state, $S_t=\left(\mathrm{SS}_t, \mathrm{PS}_t\right)$

博弈论代考

经济代写|博弈论代写Game Theory代考|MultiLayered Defense Network Structure

我们只需将 HIN、MIN 和 LIN 相应地部署到核心层、中间层和外层。具体来说，每一层由相应类型的节点组成，如下所示：

核心层由具有高重要性的节点（即 HIN）组成。由于该层中的节点具有高重要性级别，如果它们被破坏和/或失败，则很容易导致系统故障，因为高重要性节点在估计 BF 条件时被赋予更多权重，如等式 1 所示。(20.2)。示例节点可能包括云中的中央数据库服务器或身份验证服务器。HINs的重要性范围在[8-10]范围内为一个整数，每个HIN的重要性值都是根据这个范围分配的。由于该层节点的重要性很高，因此核心层中的所有节点都将成为攻击者的目标。回想一下，HIN 只能与 MIN 或 HIN 直接连接，而不能与 LIN 直接连接。该网络拓扑旨在为核心层的 HIN 增加更多的安全保护。
中间层由具有中等重要性的节点（即 MIN）组成。示例节点可能包括边缘服务器，这些边缘服务器主要将核心层中的服务器连接到外层中的传感器和物联网设备，例如路由器和交换机。MIN的重要性范围在[4-7]中设置为一个整数，每个MIN的重要性根据这个范围分配。MIN 可以直接与 HIN 或 LIN 连接，因为它们主要充当核心层和外层之间的协调器。MIN 的故障或妥协可能会在缺少 MIN 时阻碍服务可用性，此外还可能由于触发方程式中的 BF 条件而导致系统故障。(20.2)。
外层由低重要性节点（即 LIN）组成。示例节点可能包括低成本物联网设备和/或传感器。该层中的节点（即 LIN）的重要性范围为[1,3]作为一个整数。与 MIN 类似，LIN 的危害或故障可能会触发 BF 条件。LIN 可以直接与 LIN 或 MIN 相连，但不能与 HIN 相连，因为外层的 LIN 是攻击者可以访问以渗透到给定网络的第一层。缺少 LIN 可能会导致保护层较浅，这可能会导致攻击者轻松进入更深层。

为了维护这种基于优先级的多层网络结构，即使节点四处移动，它们也需要在不违反多层网络结构的情况下检查与哪些节点通信。图 20.1 显示了本节中描述的 SD-IoBT 网络结构示例。

经济代写|博弈论代写Game Theory代考|Selection of an Intrusion Response Policy

MDP 的表述：在马尔可夫决策过程 (MDP) 中，我们假设环境的完整状态信息可用，环境的下一个状态仅取决于当前状态和所选动作。
这项工作通过元组将 MDP 视为 DRL 的数学框架 $\langle\delta, \mathscr{A}, \mathscr{F}, \mathscr{R}, \gamma\rangle$ 对于 DRL 代理，由 (1) 一组状态组成 $\delta$ 具有起始状态的分布 $\mathrm{p}(\mathrm{s} 0)$ 在哪里 $\mathbf{S}t$ 指的是时间的一组状态 $t ;(2)$ 一组动作， Imathscr ${A}=\backslash$ eft $\left.\backslash \backslash m a t h b f{a} _1, \backslash d o t s, \backslash m a t h b f{a} _t, \backslash d o t s \backslash m a t h b f{a} _T \backslash r i g h t\right}$ 在哪里 $\mathbf{a}_t$ 是一组在某个时间可用的动作 $t$; (3) 转移概率， $\mathscr{T}_i\left(\mathbf{s}_l, \mathbf{a}_t, \mathbf{s} t+1\right)$ ，从状态 $\mathbf{s} l$ 陈述 $\mathbf{s} t+1$ 通过 $\mathbf{a} t$; (4) 即时奖励功能， $\mathscr{R}\left(\mathbf{s}_t, \mathbf{a}_l, \mathbf{s} l+1\right)$; (5) 奖励折扣因子 (例如 $\left.\gamma\right)$; (6) 政策 $\pi$ 从状态映射到动作的概率分布， $\pi: \delta \rightarrow(\mathscr{A}=\mathbf{a} \mid \delta) . \mathrm{RL}$ 智能体旨在最大化其自身的总预期回报 $\mathscr{R}=\sum{t=0}^T \gamma_t R_t$ 在哪里 $\gamma$ 是贴现因子，并且 $T$ 是时间范围。MDP 可以生成最优策略 $\pi^*$ 最大化来自各州的预期奖励 (Arulkumaran 等人， 2017a) .
为了选择可以最大化给定网络的安全性和性能的最佳入侵响应操作，我们通过以下方式制定 MDP:

状态 $(\delta)$ : 当时的状态 $t$ 通过采取行动，表示为 $S_l$, 是指系统安全性和性能方面的系统状态 (即 $\mathrm{SS}_l$ 和 $\mathrm{PS}_l$ ），这将在本节后面详细介绍。可以采取以下三种行动：销毁、修理或更换。第 20.4.2 节详细介绍了每个操作。与所采取的每项行动相关的奖励在本节后面阐述。特别的行动。DRL 代理将选择三个操作之一。
折扣系数 $(\gamma)$ : 我们会考虑折扣因素， $\gamma$,范围在 $(0,1]$ 作为实数。
多层网络结构的目标：在每一层，位于相应 SDN 控制器上的代理将跟踪检测为受损/失败 (NDCF) 的节点数量以及在其管理的层中正确传递的消息数量。根据每个代理通知DRL代理的信息， DRL代理将估计系统状态。为了简单起见，我们考虑系统安全状态和性能状态来表示系统状态， $S_t=\left(\mathrm{SS}_t, \mathrm{PS}_t\right)$

经济代写|博弈论代写Game Theory代考请认准statistics-lab™

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

金融工程是使用数学技术来解决金融问题。金融工程使用计算机科学、统计学、经济学和应用数学领域的工具和知识来解决当前的金融问题，以及设计新的和创新的金融产品。

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

术语广义线性模型（GLM）通常是指给定连续和/或分类预测因素的连续响应变量的常规线性回归模型。它包括多元线性回归，以及方差分析和方差分析（仅含固定效应）。

有限元方法代写

有限元方法（FEM）是一种流行的方法，用于数值解决工程和数学建模中出现的微分方程。典型的问题领域包括结构分析、传热、流体流动、质量运输和电磁势等传统领域。

有限元是一种通用的数值方法，用于解决两个或三个空间变量的偏微分方程（即一些边界值问题）。为了解决一个问题，有限元将一个大系统细分为更小、更简单的部分，称为有限元。这是通过在空间维度上的特定空间离散化来实现的，它是通过构建对象的网格来实现的：用于求解的数值域，它有有限数量的点。边界值问题的有限元方法表述最终导致一个代数方程组。该方法在域上对未知函数进行逼近。[1] 然后将模拟这些有限元的简单方程组合成一个更大的方程系统，以模拟整个问题。然后，有限元通过变化微积分使相关的误差函数最小化来逼近一个解决方案。

tatistics-lab作为专业的留学生服务机构，多年来已为美国、英国、加拿大、澳洲等留学热门地的学生提供专业的学术服务，包括但不限于Essay代写，Assignment代写，Dissertation代写，Report代写，小组作业代写，Proposal代写，Paper代写，Presentation代写，计算机作业代写，论文修改和润色，网课代做，exam代考等等。写作范围涵盖高中，本科，研究生等海外留学全阶段，辐射金融，经济学，会计学，审计学，管理学等全球99%专业科目。写作团队既有专业英语母语作者，也有海外名校硕博留学生，每位写作老师都拥有过硬的语言能力，专业的学科背景和学术写作经验。我们承诺100%原创，100%专业，100%准时，100%满意。

随机分析代写

随机微积分是数学的一个分支，对随机过程进行操作。它允许为随机过程的积分定义一个关于随机过程的一致的积分理论。这个领域是由日本数学家伊藤清在第二次世界大战期间创建并开始的。

时间序列分析代写

随机过程，是依赖于参数的一组随机变量的全体，参数通常是时间。随机变量是随机现象的数量表现，其时间序列是一组按照时间发生先后顺序进行排列的数据点序列。通常一组时间序列的时间间隔为一恒定值（如1秒，5分钟，12小时，7天，1年），因此时间序列可以作为离散时间数据进行分析处理。研究时间序列数据的意义在于现实中，往往需要研究某个事物其随时间发展变化的规律。这就需要通过研究该事物过去发展的历史记录，以得到其自身发展的规律。

回归分析代写

多元回归分析渐进（Multiple Regression Analysis Asymptotics）属于计量经济学领域，主要是一种数学上的统计分析方法，可以分析复杂情况下各影响因素的数学关系，在自然科学、社会和经济学等多个领域内应用广泛。

MATLAB代写

MATLAB 是一种用于技术计算的高性能语言。它将计算、可视化和编程集成在一个易于使用的环境中，其中问题和解决方案以熟悉的数学符号表示。典型用途包括：数学和计算算法开发建模、仿真和原型制作数据分析、探索和可视化科学和工程图形应用程序开发，包括图形用户界面构建MATLAB 是一个交互式系统，其基本数据元素是一个不需要维度的数组。这使您可以解决许多技术计算问题，尤其是那些具有矩阵和向量公式的问题，而只需用 C 或 Fortran 等标量非交互式语言编写程序所需的时间的一小部分。MATLAB 名称代表矩阵实验室。MATLAB 最初的编写目的是提供对由 LINPACK 和 EISPACK 项目开发的矩阵软件的轻松访问，这两个项目共同代表了矩阵计算软件的最新技术。MATLAB 经过多年的发展，得到了许多用户的投入。在大学环境中，它是数学、工程和科学入门和高级课程的标准教学工具。在工业领域，MATLAB 是高效研究、开发和分析的首选工具。MATLAB 具有一系列称为工具箱的特定于应用程序的解决方案。对于大多数 MATLAB 用户来说非常重要，工具箱允许您学习和应用专业技术。工具箱是 MATLAB 函数（M 文件）的综合集合，可扩展 MATLAB 环境以解决特定类别的问题。可用工具箱的领域包括信号处理、控制系统、神经网络、模糊逻辑、小波、仿真等。

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|Economic Contribution

Posted on 2023年4月6日2023年4月6日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Multiplier Concept

The multiplier concept asks for the effect of a unit investment in an industry or sector and the corresponding increase in GDP (GDP multiplier) or in employment (employment multiplier).

To understand the concept, we must look at the (macroeconomic) Keynesian consumption function $(C)$ with an autonomous consumption part $C_a$ and an income-dependent consumption $(c \times Y)$ where $c$ denotes the propensity to consume and $Y$ the income.
$$
C=C_{\mathrm{a}}+c Y
$$
In a closed economy without import and export, the national income is equivalent to GDP and can be spent on consumption $(C)$ and autonomous investments $\left(I_{\mathrm{a}}\right)$.
$$
Y=C+I_{\mathrm{a}}
$$
Replacing $C$ in Eq. (10.2) by Eq. (10.1) provides:
$$
Y=C_{\mathrm{a}}+c Y+I_{\mathrm{a}}
$$
Rearranging the equation for $Y$ yields:
$$
Y=\left(C_{\mathrm{a}}+I_{\mathrm{a}}\right) /(1-c)
$$
The derivative of Eq. (10.4) for the autonomous investment $I_{\mathrm{a}}$ provides the change in GDP $(Y):$
$$
\mathrm{d} Y / \mathrm{d} I_{\mathrm{a}}=1 / 1-c=1 / s
$$
This tells us that the multiplier $\left(\mathrm{d} Y / \mathrm{d} I_{\mathrm{a}}\right)$ is larger when the marginal propensity to consume is large or the propensity to save is small. Both values are time-dependent; in Germany, the propensity to save was $12.6 \%$ in 1991 and $9.7 \%$ in 2016.

An extension of the model can include imports and exports as well as taxes; both influence the multiplicator. Taxes $(t)$ reduce income and consequently also consumption. The multiplicator thus becomes:
$$
\mathrm{d} Y / \mathrm{d} l \mathrm{a}=1 /(1-\mathrm{c}(1-t))
$$
The propensity to import $(\mathrm{m})$ influences the multiplicator because spending for imports reduces spending at home. Considering imports, the multiplier becomes:
$$
\mathrm{d} Y / \mathrm{d} I \mathrm{a}=1 /(1-(1-t)(c-m))
$$

经济代写|博弈论代写Game Theory代考|Actors in the Construction Sector

The number of actors in the construction sector depends on the underlying definition of the sector. ISIC and NACE provide a narrow definition and the investment concept a wider one. Actors in the construction market are not only the owners as buyers and contractors as sellers as it applies to the value-added concept. Looking at the investment concept, we have to add architects, engineers, suppliers, and equipment producers. However, there are still others with important roles: administrative, legal, and professional institutions, as well as bankers and insurers. They pursue different activities during the life cycle of a built structure. Figure 10.3 is a modification of a proposal by Carassus (2004). It contains three constructs, i.e. the life cycle, the construction sector according to ISIC/NACE, and the construction sector according to the investment concept:

Construct 1 – life cycle. The life cycle consists of three main categories with more subcategories. Typically, a structure is erected and then operated or used before final demolition. New construction is usually built in months, but it can be a rather long phase for some infrastructure megaprojects; the development and design phases alone might take decades. Operation will often be by far the longest phase – one only has to think of castles or bridges that have been in use for centuries. There will be several rounds of renovation and modernization, with continuous maintenance during the operation phase. Only construction belongs to the construction sector, not facility management (operation). Demolition does not take much time.
Construct 2 – construction sector according to ISIC/NACE. This is the construction sector as defined by the value-added concept and discussed in Section 10.2.1. It includes owners and contractors as well as subcontractors. Of course, ISIC and NACE do not refer to the owner explicitly, but there would be no contract without an owner. A firm can only become a contractor by signing a construction contract with an owner.
Construct 3 – construction sector according to the investment concept. In this view, the owner receives support of architects, engineers, project managers, quantity surveyors, and other consultants. Owners might also need help from lawyers to draw up a contract, from bankers to provide financing, and from insurers to limit risks. Contractors rely on the resources of suppliers and equipment manufacturers – for which they often need the help of retailers. Like owners, contractors often require help from lawyers, insurers, and bankers. As noted in construct 3 , we find in Figure 10.3 two entities called owner and contractor:

博弈论代考

经济代写|博弈论代写Game Theory代考|Multiplier Concept

乘数既念询问单位投资对一个行业或部门的影响以及相应的 GDP 增长 (GDP 乘数) 或就业 (就业乘数) 。
要理解这个概念，我们必须看看 (宏观经济) 凯恩斯消费函数 $(C)$ 带有自主消费部分 $C_a$ 和依赖收入的消费 $(c \times Y)$ 在哪里 $c$ 表示肖费倾向和 $Y$ 收入。
$$
C=C_{\mathrm{a}}+c Y
$$
在没有进出口的封闭经济中，国民收入相当于GDP，可以用来消费 $(C)$ 和自主投资 $\left(I_{\mathrm{a}}\right)$.
$$
Y=C+I_{\mathrm{a}}
$$
更换 $C$ 在等式中 (10.2) 由等式。(10.1) 规定:
$$
Y=C_{\mathrm{a}}+c Y+I_{\mathrm{a}}
$$
重新排列方程式 $Y$ 产量:
$$
Y=\left(C_{\mathrm{a}}+I_{\mathrm{a}}\right) /(1-c)
$$
方程式的导数。(10.4) 自主投资 $I_{\mathrm{a}}$ 提供 $\mathrm{GDP}$ 的变化 $(Y):$
$$
\mathrm{d} Y / \mathrm{d} I_{\mathrm{a}}=1 / 1-c=1 / s
$$
这告诉我们乘数 $\left(\mathrm{d} Y / \mathrm{d} I_{\mathrm{a}}\right)$ 当边际消费倾向较大或储畫倾向较小时，该值较大。这两个值都与时间有关；在德国，储蓄倾向是 $12.6 \% 1991$ 年和 $9.7 \% 2016$ 年。
该模型的扩展可以包括进出口以及税收；两者都会影响乘数。税收 $(t)$ 减少收入，进而减少消费。乘数因此变为:
$$
\mathrm{d} Y / \mathrm{d} l \mathrm{a}=1 /(1-\mathrm{c}(1-t))
$$
进口倾向 $(\mathrm{m})$ 影响乘数，因为进口支出减少了国内支出。考虑进口，乘数变为:
$$
\mathrm{d} Y / \mathrm{d} I \mathrm{a}=1 /(1-(1-t)(c-m))
$$

经济代写|博弈论代写Game Theory代考|Actors in the Construction Sector

建筑行业参与者的数量取决于该行业的基本定义。ISIC 和 NACE 提供了一个狭义的定义，而投资概念则提供了一个更广泛的定义。建筑市场的参与者不仅是作为买方的业主和作为卖方的承包商，因为它适用于增值概念。从投资概念上看，我们要加上建筑师、工程师、供应商、设备生产商。然而，还有其他一些机构发挥着重要作用：行政、法律和专业机构，以及银行家和保险公司。他们在建筑结构的生命周期中从事不同的活动。图 10.3 是对 Carassus (2004) 提议的修改。它包含三个结构，即生命周期、根据 ISIC/NACE 的建筑部门、

构造 1 – 生命周期。生命周期由三个主要类别和更多子类别组成。通常，在最终拆除之前，先竖立结构然后运行或使用。新建筑通常在数月内建成，但对于一些基础设施大型项目来说，这可能是一个相当长的阶段；仅开发和设计阶段就可能需要数十年。运营通常是迄今为止最长的阶段——只需想想已经使用了几个世纪的城堡或桥梁。将进行多轮翻新和现代化改造，并在运营阶段进行持续维护。只有建筑属于建设部门，不属于设施管理（运营）部门。拆除不需要太多时间。
Construct 2 – 根据 ISIC/NACE 的建筑行业。这是增值概念所定义的建筑部门，并在第 10.2.1 节中进行了讨论。它包括业主和承包商以及分包商。当然，ISIC 和 NACE 并未明确提及所有者，但没有所有者就没有合同。公司只能通过与业主签订建筑合同才能成为承包商。
Construct 3 – 根据投资概念建设部门。从这个角度来看，业主得到了建筑师、工程师、项目经理、工料测量师和其他顾问的支持。业主可能还需要律师的帮助来起草合同，需要银行的帮助来提供融资，需要保险公司的帮助来限制风险。承包商依赖供应商和设备制造商的资源——为此他们通常需要零售商的帮助。与业主一样，承包商通常需要律师、保险公司和银行家的帮助。正如构造 3 中所指出的，我们在图 10.3 中发现了两个实体，称为所有者和承包商：

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|Multiplier Concept

Posted on 2023年4月6日2023年4月6日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Economic Contribution

We can estimate the economic contribution of the construction sector based on three concepts. The first one is the concept of value added as used for NACE and ISIC. The second one asks how much money is spent on construction. This is the investment concept, and because of the narrow definition of construction in NACE and ISIC, it is more inclusive. The third is the multiplier concept which determines how much GDP increases if a dollar is invested in construction. It is based on investments but it still encompasses more than the investment concept.

The economic contribution of construction based on the value-added concept of NACE or ISIC measures the value creation by firms engaging in the process of the construction of buildings, civil engineering projects, and specialized construction activities. Table 10.2 provides data from the Organization for Economic Co-operation and Development (OECD) on the economic contribution of different sectors to the economy in the most recent year for which data are available (Organization for Economic Co-operation and Development 2021a). As Table 10.2 clearly demonstrates, construction is an important part of the secondary sector in most countries.

Another indicator of the importance of a sector is the number of people employed. Table 10.3 provides OECD data for selected sectors (Organization for Economic Co-operation and Development 2021b).

Together, Tables 10.2 and 10.3 give ample proof that in most OECD countries the service sector is most important. It is interesting to note that construction contributes a higher percentage to employment than to value added.

To highlight the importance of construction, it might be advisable to compare different industries. I have picked three industries in Germany and assembled the data from different sources: The automobile industry – which is a global leader; the chemical industry which is the strongest in Europe; and construction (Table 10.4).

While the automobile industry is a key industry in Germany for its impact on exports, construction contributes more to GDP and employment.

经济代写|博弈论代写Game Theory代考|Investment Concept

We have seen that the NACE definition of the construction sector does not include all activities for which managers, architects, and engineers are trained and for which construction is required. Missing are:

Wood from forestry (A)
Raw materials from mining (B)
Prefabricated components from manufacturing (C)
Energy (D)
Water (E)
Architectural and engineering (M)
Renting and leasing of equipment, landscaping (N)
Public services $(\mathrm{O})$
All these activities contribute to the planning and construction of buildings and infrastructure. We are well advised to consider these activities when determining the economic impact of construction. The concept used for this purpose is construction investment, valued at the amount of money that owners spend until taking possession of a building or infrastructure. It should come as no surprise that the corresponding figures are considerably higher than those from the value-added concept.

Table 10.5 shows construction investment for several countries, ranked highest to lowest, together with the corresponding contributions to GDP in percent (European Construction Industry Federation 2021). The contributions of construction to GDP almost double when we change the framework for data collection from value added to the investment concept (EU 27: 5.7 vs. 9.5\%, shown in Table 10.2). It seems to me that the picture emerging from the investment concept better portrays the contribution of the construction sector.

Still missing from these data are all real estate activities, and this also makes sense to me because construction within the investment concept includes all activities until handover of the project to the owner (with the exception of the owner’s expenses for developing the real estate) and then other real estate activities (buying, selling, renting, operating) follow. Note that real estate together with the corresponding land value make up the largest amount of wealth in most countries.

博弈论代考

经济代写|博弈论代写Game Theory代考|Economic Contribution

我们可以根据三个概念来估算建筑业的经济贡献。第一个是用于 NACE 和 ISIC 的增值概念。第二个问建设用了多少钱。这是投资概念，由于NACE和ISIC对建筑的定义较窄，因此更具包容性。第三个是乘数概念，它决定了如果将一美元投资于建设，GDP 会增加多少。它以投资为基础，但它所包含的不仅仅是投资概念。

基于 NACE 或 ISIC 增值概念的建筑经济贡献衡量了从事建筑、土木工程项目和专业建筑活动的公司所创造的价值。表 10.2 提供了经济合作与发展组织 (OECD) 提供的最近一年不同部门对经济的经济贡献的数据（经济合作与发展组织 2021a）。正如表 10.2 清楚表明的那样，建筑业是大多数国家第二产业的重要组成部分。

一个部门重要性的另一个指标是就业人数。表 10.3 提供了选定部门的经合组织数据（经济合作与发展组织 2021b）。

表 10.2 和 10.3 一起充分证明，在大多数经合组织国家中，服务业是最重要的。有趣的是，建筑业对就业的贡献率高于对增加值的贡献率。

为了突出建筑的重要性，比较不同的行业可能是明智的。我在德国挑选了三个行业，并从不同来源收集了数据：汽车行业——全球领先；欧洲最强的化学工业；和建设（表 10.4）。

虽然汽车行业是德国影响出口的关键行业，但建筑业对 GDP 和就业的贡献更大。

经济代写|博弈论代写Game Theory代考|Investment Concept

我们已经看到，NACE 对建筑行业的定义并不包括所有管理人员、建筑师和工程师接受培训并需要施工的活动。缺少的是：

林业木材 (A)
采矿原料（B）
来自制造业的预制构件 (C)
能源 (D)
水 (E)
建筑与工程（男）
设备租赁、景观美化（N）
公共服务(欧)
所有这些活动都有助于建筑物和基础设施的规划和建设。我们强烈建议在确定建设的经济影响时考虑这些活动。用于此目的的概念是建筑投资，以所有者在拥有建筑物或基础设施之前花费的金额来衡量。相应的数字大大高于增值概念的数字也就不足为奇了。

表 10.5 显示了几个国家的建设投资，从高到低排列，以及相应的对 GDP 的贡献百分比（欧洲建筑业联合会 2021）。当我们将数据收集框架从增加值更改为投资概念时，建筑业对 GDP 的贡献几乎翻了一番（欧盟 27 国：5.7 对 9.5%，如表 10.2 所示）。在我看来，从投资概念中浮现出来的图景更好地描绘了建筑业的贡献。

这些数据中仍然缺少所有房地产活动，这对我来说也很有意义，因为投资概念中的建设包括项目移交给业主之前的所有活动（业主开发房地产的费用除外）和然后是其他房地产活动（购买、销售、租赁、经营）。请注意，在大多数国家，房地产和相应的土地价值构成了最大的财富。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|ECOS3012

Posted on 2023年3月28日2023年3月28日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Generative Adversarial Networks

GANs (Goodfellow et al. 2014), generative models first created in 2014, have achieved unprecedented success in computer vision and natural language processing. GANs consist of two competing neural networks: a generator and a discriminator. The goal of the generator is to create an example that is indistinguishable from examples in the training data. The discriminator is a critic that attempts to output 0 to the generator’s fake output and 1 to real example from the training data. The competition between these two neural networks improves each other’s performance and eventually, the generator outputs examples that are indistinguishable from the examples of the training data. We use GANs in our algorithm for two reasons: (1) GANs functionality in our objective function will strive to optimize their loss functions such that the generated adversarial examples are from the set of the original attack traffic. (2) GAN-based attack algorithm AdvGAN (Xiao et al. 2018) is known to be very effective at fooling DNN models in the image domain.

There are two main types of GAN-based algorithms to craft adversarial examples: IDSGAN (Lin et al. 2018) and AdvGAN (Xiao et al. 2018). The first algorithm IDSGAN creates adversarial examples in the network intrusion domain. The IDSGAN, depicted in Figure 17.1, first divides features of attack traffic into two groups, important features $\vec{x}1$ that preserve the semantics of the attack traffic and unimportant features $\vec{x}{\mathrm{N}}$ that do not have any effect on the semantics of the attack traffic. Then, the IDSGAN’s generator takes the unimportant features $\vec{x}{\mathrm{N}}$ as an input and replaces them with $\vec{x}{\mathrm{C}}$ without any constraints enforced. This algorithm raises two important questions: (1) Is the perturbed network traffic still valid attack traffic? (2) How do we justify the choice that divides the attack traffic’s features into unimportant and important categories? The choice of the important features in this paper is based on prior work, which is insufficient for two reasons. First, the important features of a specific attacks might vary from one environment to a different environment. Second, categorizing attack traffic’s features into important and unimportant groups is inherently complex.

The AdvGAN (Figure 17.2) is the first algorithm that crafts adversarial examples by applying perturbations to the original instance. The algorithm outperformed many popular attacks such as PGD, C\&W, FGSM, and JSMA in blackbox settings and won first place in Madry’s competition held by MIT. However, the algorithm is created for only image domain. We modified this algorithm such that the crafted adversarial example obeys the constraints of the network traffic.

The AdvGAN’s goal is to train a generator that crafts a perturbation $\Delta x$ for an original instance $x$ so that the classification of the $x+\Delta x$ by a target model becomes the class the adversary wants.

经济代写|博弈论代写Game Theory代考|Target NIDS for Detecting Attack Traffic

ML-based NIDS are often trained with a database consisting of both attack and benign traffic. In this work, we evaluated our attack algorithm against two state-of-the-art DNN-based NIDS models as well as our own NIDS model.

To find the optimal DNN model for the NIDS, we vary the number of layers and the number of neurons in the layers depending on the dataset. Theoretically, the optimal number of hidden layers of the DNN is related to the complexity of the datasets. In this work, we choose an optimal architecture of a target NIDS for each dataset based on the evaluations as shown in Tables 17.1 and 17.2. We experimented with five different DNNs differentiated by their number of layers. Each layer of DNNs consists of a number of perceptrons, a ReLU activation function, and a dropout function that prevents overfitting. For each case, we calculate their accuracy, precision, recall, and $F$-score as shown in Tables 17.3 and 17.4 (where one-layer DNN represents a multiclass logistic regression). For choosing the number of perceptrons in the hidden layers, we used the benchmark rule that the number of perceptrons in the hidden layers should be between the number of neurons in input and output layer (the number of perceptrons are modified for performance improvement). After a number of iterations, we find that the following architectures can provide results as high as the state-of-the-art results shown in the previous work (Abdulraheem and Ibraheem 2019; Gao et al. 2020; Panigrahi and Borah 2018; Rahul Vigneswaran et al. 2018). As shown from the experiments in Tables 17.3 and 17.4, adding more layers to three-layer DNN does not improve the accuracy due to overfitting, but it increases the computational cost.

For these reasons, we choose three-layer DNN as our target NIDS. The prediction accuracy, architecture, and hyperparameters of the target models’ are shown in Table 17.5. According to published results of the DNN-models trained on CICIDS and NSL-KDD datasets, the prediction accuracies we found in our models are at the same level as the state-of-the-art predictions (Vinayakumar et al. 2019). In the next section, we introduce domain constraints.

博弈论代考

经济代写|博弈论代写Game Theory代考|Generative Adversarial Networks

GAN（Goodfellow 等人，2014 年）是 2014 年首次创建的生成模型，在计算机视觉和自然语言处理方面取得了前所未有的成功。GAN 由两个相互竞争的神经网络组成：生成器和鉴别器。生成器的目标是创建一个与训练数据中的示例无法区分的示例。鉴别器是一个批评者，它试图将 0 输出到生成器的假输出，将 1 输出到训练数据中的真实示例。这两个神经网络之间的竞争提高了彼此的性能，最终生成器输出的示例与训练数据的示例无法区分。我们在算法中使用 GAN 有两个原因：(1) 我们目标函数中的 GAN 功能将努力优化其损失函数，以便生成的对抗样本来自原始攻击流量的集合。(2) 基于 GAN 的攻击算法 AdvGAN (Xiao et al. 2018) 被认为在图像域中非常有效地欺骗 DNN 模型。

有两种主要类型的基于 GAN 的算法来制作对抗样本：IDSGAN (Lin et al. 2018) 和 AdvGAN (Xiao et al. 2018)。第一个算法 IDSGAN 在网络入侵域中创建对抗样本。如图 17.1 所示，IDSGAN 首先将攻击流量的特征分为两组，重要特征X→1保留攻击流量的语义和不重要的特征X→否对攻击流量的语义没有任何影响。然后，IDSGAN 的生成器采用不重要的特征X→否作为输入并将它们替换为X→C没有强制执行任何约束。该算法提出了两个重要问题：(1) 扰动的网络流量是否仍然是有效的攻击流量？(2) 我们如何证明将攻击流量的特征分为不重要和重要类别的选择？本文中重要特征的选择是基于先前的工作，这有两个原因是不够的。首先，特定攻击的重要特征可能因环境而异。其次，将攻击流量的特征分为重要组和不重要组本身就很复杂。

AdvGAN（图 17.2）是第一个通过对原始实例应用扰动来制作对抗样本的算法。该算法在黑盒设置中优于PGD、C\&W、FGSM、JSMA等多种流行攻击，在麻省理工学院举办的Madry’s竞赛中获得第一名。但是，该算法仅针对图像域创建。我们修改了该算法，使精心制作的对抗性示例遵守网络流量的约束。

AdvGAN 的目标是训练一个生成扰动的生成器丁X对于原始实例X这样分类的X+丁X通过目标模型成为对手想要的类别。

经济代写|博弈论代写Game Theory代考|Target NIDS for Detecting Attack Traffic

基于 ML 的 NIDS 通常使用包含攻击和良性流量的数据库进行训练。在这项工作中，我们针对两个最先进的基于 DNN 的 NIDS 模型以及我们自己的 NIDS 模型评估了我们的攻击算法。

为了找到适用于 NIDS 的最佳 DNN 模型，我们根据数据集改变层数和层中神经元的数量。理论上，DNN 的最佳隐藏层数与数据集的复杂性有关。在这项工作中，我们根据表 17.1 和 17.2 中所示的评估为每个数据集选择目标 NIDS 的最佳架构。我们试验了五种不同的 DNN，这些 DNN 的层数有所不同。DNN 的每一层都由多个感知器、一个 ReLU 激活函数和一个防止过拟合的 dropout 函数组成。对于每个案例，我们计算它们的准确度、精确度、召回率和F-score 如表 17.3 和 17.4 所示（其中一层 DNN 表示多类逻辑回归）。为了选择隐藏层中的感知器数量，我们使用了基准规则，即隐藏层中的感知器数量应介于输入和输出层中的神经元数量之间（为了提高性能而修改了感知器的数量）。经过多次迭代，我们发现以下架构可以提供与之前工作中显示的最先进结果一样高的结果（Abdulraheem 和 Ibraheem 2019；Gao 等人 2020；Panigrahi 和 Borah 2018；Rahul Vigneswaran 等人，2018 年）。从表 17.3 和 17.4 的实验可以看出，在三层 DNN 中增加更多的层并没有因为过拟合而提高准确率，反而增加了计算成本。

由于这些原因，我们选择三层 DNN 作为我们的目标 NIDS。目标模型的预测精度、架构和超参数如表 17.5 所示。根据在 CICIDS 和 NSL-KDD 数据集上训练的 DNN 模型的已发布结果，我们在模型中发现的预测精度与最先进的预测处于同一水平（Vinayakumar 等人，2019 年）。在下一节中，我们将介绍域约束。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|ECON6025

Posted on 2023年3月28日2023年3月28日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Attribution-based Confidence Using Shapley Values

While improving the accuracy and robustness of DNNs has received significant attention, there is an urgent need to quantitatively characterize the limitations of these models and improve the transparency of their failure modes. We need to define an easy to compute confidence metric on predictions of a DNN that closely reflects its accuracy. This can be used to detect adversarial examples as low confidence predictions of the model. Our approach for computing attribution-based confidence used Shapley values for computing importance of input features. Shapley values often employ the notion of a baseline input $x^b$; for example, the all dark image can be the baseline for images. The baseline can also be a set of random inputs where attribution is computed as an expected value. Let the attribution for the $j$-th feature and output label $i$ be $\mathscr{A l}_j^i(\mathbf{x})$. The attribution for the $j$-th input feature depends on the complete input $x$ and not just $x_j$. The treatment for each logit is similar, and so, we drop the logit/class and denote the network output simply as $\mathscr{F}(\cdot)$ and attribution as $\mathscr{A}_j(\mathbf{x})$. For simplicity, we use the baseline input $\mathbf{x}^b=0$ for computing attributions. We make the following two assumptions on the DNN model and the attributions, which reflect the fact that the model is well-trained and the attribution method is well-founded:

The attribution is dominated by the linear term. This is also an assumption made by attribution methods based on Shapley values such as Integrated Gradient (Sundararajan et al. 2017) which define attribution as the path integral of the gradients of the DNN output with respect to that feature along the path from the baseline $\mathbf{x}^b$ to the input $\mathbf{x}$, that is,
$$
\mathscr{A}j^l(\mathbf{x})=\left(\mathbf{x}_j-\mathbf{x}_j^b\right) \times \int{a=0}^1 \partial_j \mathscr{F}^l\left(\mathbf{x}^b+\alpha\left(\mathbf{x}-\mathbf{x}^b\right)\right) \mathrm{d} \alpha
$$
where the gradient of $i$-th logit output of the model along the $j$-th feature is denoted by $\partial_j \mathscr{F}^i(\cdot)$.
Attributions are complete i.e. the following is true for any input $\mathbf{x}$ and the baseline input $\mathbf{x}^b$ :
$\mathscr{F}(\mathbf{x})-\mathscr{F}\left(\mathbf{x}^b\right)=\sum_{k=1}^n \mathscr{A}_k(\mathbf{x})$ where $\mathbf{x}$ has $n$ features.
Shapley value methods such as Integrated Gradient and DeepShap (Sundararajan et al. 2017; Lundberg and Lee 2017) satisfy this axiom too.

The details of the attribution-based approach to compute confidence is presented in Jha et al. (2019).

经济代写|博弈论代写Game Theory代考|Network Intrusion Detection Systems

A network intrusion detection system (NIDS) is software that detects malicious activities that are intended to steal or corrupt sensitive information in computer networks. The NIDS have two main types: anomaly-based and misuse-based. In this chapter, we focus on the misuse-based detection systems. The misuse-based NIDS are divided into two main categories: knowledge-based and ML-based. In knowledge-based NIDS (also known as a signature-based detection system), attack network traffic flows are directly compared with previously defined rules or attack patterns. One of the drawbacks of signature-based NIDS is that they are not able to detect mutations of known attacks. ML-based misuse detection systems, on the other hand, learn from database signatures and predict the possible mutations of known attacks, therefore, they are widely used for academic and commercial purposes. In this chapter, we will specifically focus on ML-based misuse NIDS.

Deep neural networks (DNNs) can easily be manipulated by adversarial examples (carefully crafted inputs). In 2013, Szegedy et al.(Papernot et al. 2016), and Biggio et al. (2013), first discovered the existence of adversarial examples in the image domain. By applying a small perturbation unnoticeable to human eyes to an original instance of an image, they were able to change the classification of a trained model for the image. By exploiting this vulnerability of neural networks, adversaries can potentially manipulate self-driving cars, smart devices, and intrusion detection systems that rely on DNNs. Given an input $\vec{x}$ and a trained target model $F(\vec{x})$, an adversary tries to find the minimum perturbation $\Delta \vec{x}$ under some norm $\left(l^0, l^2\right.$, or $\left.l^{\infty}\right)$ (Carlini and Wagner 2017) to cause the $F(\vec{x}+\Delta \vec{x})$ to be classified as a specific target label $t \neq F(\vec{x})$ (Szegedy et al. 2013). This can generally be expressed by an optimization problem where $c$ is a constant that is found by binary search and $t$ is the target class (Carlini and Wagner 2017; Szegedy et al. 2013).
$$
\begin{aligned}
& \min _{\Delta \vec{x}} c|\Delta \vec{x}|+\operatorname{Loss}(F(\vec{x}+\Delta \vec{x}), t) \
& \text { such that }(\vec{x}+\Delta \vec{x}) \in[0,1]^m
\end{aligned}
$$
where $m$ is the dimension of the vector $\vec{x}$. As mentioned, most of the work has been done in unconstrained domains, specifically in the image domain. However, most of the domains in the real world are constrained. In this work, we create adversarial examples in a network intrusion detection domain where we determine and enforce the domain constraints. In our problem, $F$ is the target ML-based NIDS we want to fool and the target class $t$ is the benign class.

博弈论代考

经济代写|博弈论代写Game Theory代考|Attribution-based Confidence Using Shapley Values

虽然提高 DNN 的准确性和鲁棒性受到了极大的关注，但迫切需要定量表征这些模型的局限性并提高其故障模式的透明度。我们需要为 DNN 的预测定义一个易于计算的置信度度量，以密切反映其准确性。这可用于将对抗性示例检测为模型的低置信度预测。我们用于计算基于属性的置信度的方法使用 Shapley 值来计算输入特征的重要性。Shapley 值通常采用基线输入的概念 $x^b$; 例如，全暗图像可以作为图像的基线。基线也可以是一组随机输入，其中属性被计算为预期值。让归因于 $j$-th 特征和输出标签 $i$ 是 $\mathscr{A} \mathrm{l}_j^i(\mathbf{x})$. 归因于 $j$-th 输入特征取决于完整的输入 $x$ 而不仅仅是 $x_j$. 每个 logit 的处理都是相似的，因此，我们删除 logit/class 并将网络输出简单地表示为 $\mathscr{F}(\cdot)$ 并归因于 $\mathscr{A}_j(\mathbf{x})$. 为简单起见，我们使用基线输入 $\mathbf{x}^b=0$ 用于计算归因。我们对 DNN 模型和归因做出以下两个假设，这反映了模型训练良好和归因方法有根据的事实:

归因由线性项支配。这也是基于 Shapley 值的归因方法所做的假设，例如 Integrated Gradient (Sundararajan et al. 2017)，它将归因定义为 DNN 输出的梯度相对于沿基线路径的特征的路径积分 $\mathbf{x}^b$ 到输入 $\mathbf{x}$ ，那是，
$$
\mathscr{A} j^l(\mathbf{x})=\left(\mathbf{x}_j-\mathbf{x}_j^b\right) \times \int a=0^1 \partial_j \mathscr{F}^l\left(\mathbf{x}^b+\alpha\left(\mathbf{x}-\mathbf{x}^b\right)\right) \mathrm{d} \alpha
$$
其中的梯度 $i$-模型的第 logit 输出 $j$-th 特征表示为 $\partial_j \mathscr{F}^i(\cdot)$.
归因是完整的，即以下内容对于任何输入都是正确的 $\mathbf{x}$ 和基线输入 $\mathbf{x}^b$ : $\mathscr{F}(\mathbf{x})-\mathscr{F}\left(\mathbf{x}^b\right)=\sum_{k=1}^n \mathscr{A}_k(\mathbf{x})$ 在哪里 $\mathbf{x}$ 有 $n$ 特征。
Shapley 值方法，如 Integrated Gradient 和 DeepShap (Sundararajan et al. 2017; Lundberg and Lee 2017) 也满足这个公理。
Jha 等人介绍了基于属性的计算置信度方法的详细信息。（2019）。

经济代写|博弈论代写Game Theory代考|Network Intrusion Detection Systems

网络入侵检测系统 (NIDS) 是检测旨在窈取或破坏计算机网络中敏感信息的恶意活动的软件。NIDS 有两种主要类型：基于异常的和基于滥用的。在本章中，我们重点关注基于误用的检测系统。基于滥用的 NIDS 分为两大类：基于知识的和基于 ML 的。在基于知识的 NIDS（也称为基于签名的检测系统) 中，攻击网络流量直接与先前定义的规则或攻击模式进行比较。基于签名的 NIDS 的缺点之一是它们无法检测已知攻击的突变。另一方面，基于 $M L$ 的滥用检测系统从数据库签名中学习并预测已知攻击的可能突变，因此，它们被广泛用于学术和商业目的。在本章中，我们将特别关注基于 ML 的误用 NIDS。
深度神经网络 (DNN) 可以很容易地被对抗性示例（精心设计的输入) 操纵。2013 年，Szegedy 等人 (Papernot 等人，2016 年) 和 Biggio 等人。(2013)，首先发现图像域中存在对抗样本。通过对图像的原始实例应用人眼无法察觉的小扰动，他们能够改变图像训练模型的分类。通过利用神经网络的这一漏洞，对手可能会操纵依赖 DNN 的自动驾驶汽车、智能设备和入侵检测系统。给定一个输入䒣和训练有素的目标模型 $F(\vec{x})$ ，对手试图找到最小扰动 $\Delta \vec{x}$ 在某种规范下 $\left(l^0, l^2\right.$ ，或者 $\left.l^{\infty}\right)$ (Carlini 和 Wagner 2017) 导致 $F(\vec{x}+\Delta \vec{x})$ 被归类为特定目标标签 $t \neq F(\vec{x})$ (Szegedy 等人，2013 年)。这通常可以用优化问题表示，其中 $c$ 是通过二分查找找到的常数，并且 $t$ 是目标类别（Carlini 和 Wagner 2017；Szegedy 等人 2013) .
$$
\min _{\Delta \vec{x}} c|\Delta \vec{x}|+\operatorname{Loss}(F(\vec{x}+\Delta \vec{x}), t) \quad \text { such that }(\vec{x}+\Delta \vec{x}) \in[0,1]^m
$$
在哪里 $m$ 是向量的维度 $\vec{x}$. 如前所述，大部分工作都是在不受约束的领域完成的，特别是在图像领域。然而，现实世界中的大多数域都是受限的。在这项工作中，我们在网络入侵检测域中创建对抗性示例，我们在其中确定并执行域约束。在我们的问题中， $F$ 是我们想要欺骗的目标基于 ML 的 NIDS 和目标类 $t$ 是良性类。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|ECON3503

Posted on 2023年3月24日2023年3月24日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Further Discussions

Game theory attempts to bring mathematical precision to decision-making so that best strategies can be played, even in the arms race between the learner and the adversary. In game theory, we assume the players are rational-they seek to maximize their payoffs or minimize their losses. This assumption is not necessarily valid in real life, especially in cybersecurity domains. What appears irrational to one player may be rational to the opponent player. For example, when we model a game between airport security and terrorists, it may not be a good idea to view the terrorists based on our own experiences. When the opponent is not rational or simply plays poorly, it is important to realize that continuing to play the equilibrium strategy will lead to a losing situation or the loss of opportunity to exploit the opponent’s weaknesses.

Behavioral economists have long questioned the strict assumptions about rationality in existing theories in game theory (Aumann 1997). The actual decision-making by individuals is often irrational even in simple decision problems. Individuals typically fail to select the best response from a set of choice. As a matter of fact, optimization is so difficult that individuals are often unable to come up with their best responses to their opponents’ moves, unlike what the rational decision theory has always assumed. In response, theories that relax the rationality constraints have been proposed among which include: Quantal Response Equilibrium (QRE) (McKelvey and Palfrey 1995), a solution concept that promotes an equilibrium notion with bounded rationality that acknowledges the possibility that players do not always play a best response; Prospect Theory (Kahneman and Tversky 1979; Tversky and Kahneman 1992), a solution that introduces payoffs with respect to a reference point at which there is larger slope for losses than for gains and agents tend to overestimate small probabilities and underestimate large probabilities.

The same debate is applicable to mixed strategy games. In the airport security game, how the airport deploys security patrols depends on the response from the terrorists. If the rationality of the terrorists is predictable, it would be best for the airport to deploy its security patrols randomly by playing a mixed strategy. However, when there is a good reason to believe that the terrorists are not going to play the equilibrium strategy, airport security may be better off by playing pure strategies. Generally speaking, unless the odds are strongly in our favor, playing equilibrium strategies would be our best choice. Nevertheless, we should always keep in mind that our calculation of rationality may lead to different behavior, new rules may need to be defined in the game to take this into consideration.

经济代写|博弈论代写Game Theory代考|Adversarial Machine Learning

While there is a growing interest in applying machine learning to different data domains and deploying machine learning algorithms in real systems, it has become imperative to understand vulnerabilities of machine learning in the presence of adversaries. To that end, adversarial machine learning (Kurakin et al. 2016a; Vorobeychik and Kantarcioglu 2018; Shi et al. 2018b) has emerged as a critical field to enable safe adoption of machine learning subject to adversarial effects. One example that has attracted recent attention involves machine learning applications offered to public or paid subscribers via APIs; e.g. Google Cloud Vision (2020) provides cloud-based machine learning tools to build machine learning models. This online service paradigm creates security concerns of adversarial inputs to different machine learning algorithms ranging from computer vision to NLP (Shi et al. 2018c,d). As another application domain, automatic speech recognition and voice controllable systems were studied in terms of the vulnerabilities of their underlying machine learning algorithms (Vaidya et al. 2016; Zhang et al. 2017). As an effort to identify vulnerabilities in autonomous driving, attacks on self-driving vehicles were demonstrated in Kurakin et al. (2016), where the adversary manipulated traffic signs to confuse the learning model.

The manipulation in adversarial machine learning may happen during the training or inference (test) time, or both. During the training time, the goal of the adversary is to provide wrong inputs (features and/or labels) to the training data such that the machine learning algorithm is not properly trained. During the test time, the goal of the adversary is to provide wrong inputs (features) to the machine algorithm such that it returns wrong outputs. As illustrated in Figure 14.1, attacks built upon adversarial machine learning can be categorized as follows.

Attack during the test time.
a. Inference (exploratory) attack: The adversary aims to infer the machine learning architecture of the target system to build a shadow or surrogate model that has the same functionality as the original machine learning architecture (Barreno et al. 2006; Tramer et al. 2016; Wu et al. 2016; Papernot et al. 2017; Shi et al. 2017; Shi et al. 2018b). This corresponds to a white-box or black-box attack depending on whether the machine learning model such as the deep neural network structure is available to the adversary, or not. For a black-box attack, the adversary queries the target classifier with a number of samples and records the labels. Then, it uses this labeled data as its own training data to train a functionally equivalent (i.e. statistically similar) deep learning classifier, namely a surrogate model. Once the machine learning functionality is learned, the adversary can use the inference results obtained from the surrogate model for subsequent attacks such as confidence reduction or targeted misclassification.
b. Membership inference attack: The adversary aims to determine if a given data sample is a member of the training data, i.e. if a given data sample has been used to train the machine learning algorithm of interest (Nasr et al. 2018; Song et al. 2018; Jia et al. 2019; Leino and Fredrikson 2020). Membership inference attack is based on the analysis of overfitting to check whether a machine learning algorithm is trained for a particular data type, e.g. a particular type of images. By knowing which type of data the machine learning algorithm is trained to classify, the adversary can then design a subsequent attack more successfully.

博弈论代考

经济代写|博弈论代写Game Theory代考|Further Discussions

博弈论试图将数学精确性带入决策制定，以便即使在学习者和对手之间的军备竞赛中也能发挥最佳策略。在博弈论中，我们假设参与者是理性的——他们寻求最大化收益或最小化损失。这种假设在现实生活中不一定有效，尤其是在网络安全领域。对一个玩家来说不合理的事情可能对对手玩家来说是合理的。例如，当我们模拟机场安全和恐怖分子之间的博弈时，根据我们自己的经验来看待恐怖分子可能不是一个好主意。当对手不理性或只是打得不好时，重要的是要意识到继续采用均衡策略将导致失败的局面或失去利用对手弱点的机会。

长期以来，行为经济学家一直质疑博弈论中现有理论对理性的严格假设（Aumann 1997）。即使在简单的决策问题中，个人的实际决策也往往是非理性的。个人通常无法从一组选择中选择最佳反应。事实上，优化是如此困难，以至于个人往往无法对对手的举动做出最佳反应，这与理性决策理论一直假设的不同。作为回应，人们提出了放松理性约束的理论，其中包括：量子响应平衡 (QRE)（McKelvey 和 Palfrey 1995），这是一种解决方案概念，它促进具有有限理性的均衡概念，承认玩家并不总是玩游戏的可能性最好的回应；

同样的争论也适用于混合策略游戏。在机场安保博弈中，机场如何部署安保巡逻，取决于恐怖分子的反应。如果恐怖分子的理性是可以预见的，机场安保巡逻最好是随机部署，玩混合策略。然而，当有充分的理由相信恐怖分子不会采用均衡策略时，机场安全可能会通过采用纯策略而变得更好。一般而言，除非赔率对我们有利，否则采用均衡策略将是我们的最佳选择。然而，我们应该始终牢记，我们的理性计算可能会导致不同的行为，可能需要在游戏中定义新的规则来考虑到这一点。

经济代写|博弈论代写Game Theory代考|Adversarial Machine Learning

尽管人们对将机器学习应用于不同的数据域并在真实系统中部署机器学习算法的兴趣越来越大，但了解机器学习在存在对手的情况下的弱点已变得势在必行。为此，对抗性机器学习（Kurakin 等人 2016a；Vorobeychik 和 Kantarcioglu 2018 年；Shi 等人 2018b）已成为一个关键领域，可以安全采用受对抗影响的机器学习。最近引起关注的一个例子涉及通过 API 向公众或付费用户提供的机器学习应用程序；例如 Google Cloud Vision (2020) 提供基于云的机器学习工具来构建机器学习模型。这种在线服务范式对从计算机视觉到 NLP 的不同机器学习算法的对抗性输入产生了安全问题（Shi 等人，2018c，d）。作为另一个应用领域，自动语音识别和语音可控系统根据其底层机器学习算法的漏洞进行了研究（Vaidya 等人 2016 年；Zhang 等人 2017 年）。为了识别自动驾驶中的漏洞，Kurakin 等人演示了对自动驾驶车辆的攻击。(2016)，对手操纵交通标志来混淆学习模型。2016; 张等。2017）。为了识别自动驾驶中的漏洞，Kurakin 等人演示了对自动驾驶车辆的攻击。(2016)，对手操纵交通标志来混淆学习模型。2016; 张等。2017）。为了识别自动驾驶中的漏洞，Kurakin 等人演示了对自动驾驶车辆的攻击。(2016)，对手操纵交通标志来混淆学习模型。

对抗性机器学习中的操作可能发生在训练或推理（测试）时间，或两者兼而有之。在训练期间，对手的目标是向训练数据提供错误的输入（特征和/或标签），从而导致机器学习算法无法正确训练。在测试期间，对手的目标是向机器算法提供错误的输入（特征），使其返回错误的输出。如图 14.1 所示，基于对抗性机器学习的攻击可分为以下几类。

在测试时间内攻击。
A。推理（探索性）攻击：对手旨在推断目标系统的机器学习架构，以构建与原始机器学习架构具有相同功能的影子或替代模型（Barreno 等人，2006 年；Tramer 等人，2016 年； Wu 等人 2016 年；Papernot 等人 2017 年；Shi 等人 2017 年；Shi 等人 2018b)。这对应于白盒或黑盒攻击，具体取决于对手是否可以使用深度神经网络结构等机器学习模型。对于黑盒攻击，对手使用大量样本查询目标分类器并记录标签。然后，它使用这个标记数据作为自己的训练数据来训练一个功能等效（即统计相似）的深度学习分类器，即代理模型。
b. 成员推理攻击：对手旨在确定给定数据样本是否是训练数据的成员，即给定数据样本是否已用于训练感兴趣的机器学习算法（Nasr 等人，2018 年；Song 等人，2018 年）。 2018 年；Jia 等人 2019 年；莱诺和弗雷德里克森 2020 年）。Membership inference attack 是基于过度拟合的分析，以检查机器学习算法是否针对特定数据类型（例如特定类型的图像）进行了训练。通过了解训练机器学习算法分类的数据类型，对手可以更成功地设计后续攻击。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|ECON90022

Posted on 2023年3月24日2023年3月24日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|A Single Leader Multi-followers Stackelberg Game

In a single leader multiple followers (SLMF) game (Basar and Olsder 1999), the leader makes its optimal decision prior to the decisions of multiple followers. The Stackelberg game played by the leader is:
$$
\begin{array}{ll}
\min _{s^{\ell}, s^} & F\left(s^{\ell}, s^\right) \
\text { s.t. } & G\left(s^{\ell}, s^\right) \leq 0 \ & H\left(s^{\ell}, s^\right)=0
\end{array}
$$
where $F$ is the leader’s objective function, constrained by $G$ and $H ; s^{\prime}$ is the leader’s decision and $s^$ is in the set of the optimal solutions of the lower level problem: $$ s^ \in\left{\begin{array}{ll}
\underset{s_i}{\operatorname{argmin}} & f_i\left(s^t, s_i\right) \
\text { s.t. } & g_i\left(s^{\ell}, s_i\right) \leq 0 \
& h_i\left(s^{\ell}, s_i\right)=0
\end{array}\right} \quad \forall i=1, \ldots, m
$$
where $m$ is the number of followers, $f_i$ is the $i^{\text {th }}$ follower’s objective function constrained by $g_l$ and $h_i$

For the sake of simplicity, we assume the followers are not competing among themselves. This is usually a valid assumption in practice since adversaries rarely affect each other through their actions. In a Bayesian Stackelberg game, the followers may have many different types and the leader does not know exactly the types of adversaries it may face when solving its optimization problem. However, the distribution of the types of adversaries is known or can be inferred from past experience. The followers’ strategies and payoffs are determined by the followers’ types. The followers play their optimal responses to maximize the payoffs given the leader’s strategy. The Stackelberg equilibrium includes an optimal mixed strategy of the learner and corresponding optimal strategies of the followers.

Problem Definition Given the payoff matrices $R^{\ell}$ and $R^f$ of the leader and the $m$ followers of $n$ different types, find the leader’s optimal mixed strategy given that all followers know the leader’s strategy when optimizing their rewards. The leader’s pure strategies consist of a set of generalized linear learning models $\langle\phi(x), w\rangle$ and the followers’ pure strategies include a set of vectors performing data transformation $x \rightarrow x+\delta x$.

Given the payoff matrices $R^{\ell}$ and $R^f$ discussed in Section 13.5.3.3 (as shown in Table 13.6), let $r^f$ denote the follower’s maximum payoff, $\mathcal{L}$ and $\mathcal{F}$ denote the indices of the pure strategies in the leader’s policy $s^{\ell}$ and the follower’s policy $s^f$.

经济代写|博弈论代写Game Theory代考|Real Datasets

For the two real datasets we again use spam base (UCI Machine Learning Repository 2019) and web spam (LIBSVM Data 2019). The learning tasks are binary classification problems, differentiating spam or webspam from legitimate e-mail or websites.

Spambase Dataset Recall that in the spam base dataset, the task is to differentiate spam from legitimate e-mail. There are 4601 e-mail messages in the dataset including approximately 1800 spam messages. The dataset has 57 attributes and one class label. Training and test datasets are separate. The results are averaged over 10 random runs. The detailed results are shown in Table 13.8. The $f_{\mathrm{a}}=0$ column is left out for the same reason as explained earlier on the artificial datasets.

This dataset serves as an excellent example to demonstrate the power of the mixed strategy. In the cases where $p=0.1$, that is, when we assume legitimate e-mail is modified $10 \%$ of the time (while spam is always modified), the two equilibrium predictors Equi $^{* 1}$ and Equi ${ }^{* 2}$ exhibit very stable performance in terms of predictive accuracy. Their error rates fluctuate slightly at 0.37 regardless of how aggressively the test data has been modified. On the other hand, $S \mathrm{SM}^{* 1}$ and $S_{V M}{ }^{* 2}$ significantly outperform the equilibrium predictors Equi ${ }^{* 1}$ and Equi $^{* 2}$ and the invariant SVM when $f_{\mathrm{a}} \leq 0.5$. However, the performance of SVM ${ }^{* 1}$ and SVM ${ }^{* 2}$ dropped quickly as the attack gets more intense $\left(f_{\mathrm{a}}>0.5\right)$, much poorer than the equilibrium predictors. The mixed strategy, although not the best, demonstrates superb performance by agreeing with the winning models the majority of the time. The standard SVM has similar performance to the equilibrium predictors, behaving poorly as the attack gets intense. When $p=0.5$, that is, when legitimate e-mail is modified half of the time while all spam is modified, equilibrium predictors still demonstrate very stable performance while the performance of the equilibrium predictors Equi $^{* 1}$ and Equi ${ }^{* 2}$ deteriorates sharply right after the attack factor increases to 0.3 . The mixed strategy, again not the best predictor, demonstrates the most consistent performance among all the predictors given any attack intensity levels.

We also tested the case where the attack factor $f_{\mathrm{a}} \in(0,1)$ is completely random under uniform distribution for each attacked sample on this dataset. The probability of negative data being attacked increases gradually from 0.1 to 0.9 . The results are illustrated in Figure 13.11. Again, we observe similar behavior of all the predictors: stable equilibrium predictors, $\mathrm{SVM}^{* 1}, \mathrm{SVM}^{* 2}$, and SVM progressively deteriorating as $p$ increases. The mixed strategy, although weakened as more negative data is allowed to be modified, consistently lies in between the equilibrium predictors and the $S V M$ predictors.

博弈论代考

经济代写|博弈论代写Game Theory代考|A Single Leader Multi-followers Stackelberg Game

在单一领导者多追随者 (SLMF) 博亦中（Basar 和 Olsder 1999），领导者在多个追随者做出决定之前做出其最优决策。领导者玩的 Stackelberg 游戏是:
在哪里 $F$ 是领导者的目标函数，受制于 $G$ 和 $H ; s^{\prime}$ 是领导者的决定^在较低级别问题的最优解集合中:
在哪里 $m$ 是追随者的数量， $f_i$ 是个 $i^{\text {th }}$ 追随者的目标函数受制于 $g_l$ 和 $h_i$
为了简单起见，我们假设追随者之间没有竞争。这在实践中通常是一个有效的假设，因为对手很少通过他们的行动相互影响。在贝叶斯 Stackelberg 博恋中，追随者可能有许多不同的类型，领导者在解决其优化问题时并不知道它可能面临的对手的确切类型。然而，对手类型的分布是已知的，或者可以从过去的经验中推断出来。追随者的策略和收益由追随者的类型决定。追随者根据领导者的策略发挥他们的最佳反应以最大化收益。Stackelberg 均衡包括学习者的最优混合策略和跟随者的相应最优策略。
问题定义给定支付矩阵 $R^{\ell}$ 和 $R^f$ 领导者和 $m$ 追随者 $n$ 不同的类型，找到领导者的最优混合策略，因为所有追随者在优化他们的奖励时都知道领导者的策略。领导者的纯策略由一组广义线性学习模型组成 $\langle\phi(x), w\rangle$ 跟随者的纯策略包括一组执行数据转换的向量 $x \rightarrow x+\delta x$.
给定支付矩阵 $R^{\ell}$ 和 $R^f$ 在第 13.5.3.3 节中讨论过 (如表 13.6 所示)，让 $r^f$ 表示跟随者的最大收益， $\mathcal{L}$ 和 $\mathcal{F}$ 表示领导者政策中纯策略的指标 $s^{\ell}$ 和追随者的政策 $s^f$.

经济代写|博弈论代写Game Theory代考|Real Datasets

对于这两个真实数据集，我们再次使用垃圾邮件库 (UCI Machine Learning Repository 2019) 和网络垃圾邮件 (LIBSVM Data 2019)。学习任务是二元分类问题，将垃圾邮件或网络垃圾邮件与合法电子邮件或网站区分开来。

Spambase 数据集回想一下，在垃圾邮件基础数据集中，任务是区分垃圾邮件和合法电子邮件。数据集中有 4601 封电子邮件，包括大约 1800 封垃圾邮件。该数据集有 57 个属性和一个类标签。训练和测试数据集是分开的。结果取 10 次随机运行的平均值。详细结果如表13.8所示。这 $f_{\mathrm{a}}=0$ 出于与前面在人工数据集上解释的相同原因，列被遗漏了。
该数据集是展示混合策略强大功能的绝佳示例。在这种情况下 $p=0.1$ ，也就是说，当我们假设合法的电子邮件被修改时 $10 \%$ 的时间 (而垃圾邮件总是被修改)，两个均衡预测器 Equi1 和装备 ${ }^{ 2}$ 在预测准确性方面表现出非常稳定的性能。无论测试数据修改得多么激进，它们的错误率都在 0.37 处轻微波动。另一方面， $S \mathrm{SM}^{* 1}$ 和 $S_{V M}{ }^{* 2}$ 显着优于均衡预测变量 Equi1 和装备 ${ }^{ 2}$ 和不变的 SVM 当 $f_{\mathrm{a}} \leq 0.5$. 但是，SVM 的性能1 和支持向量机 ${ }^{ 2}$ 随着攻击变得更加激烈而迅速下降 $\left(f_{\mathrm{a}}>0.5\right)$ ，比均衡预测变量差得多。混合策略虽然不是最好的，但在大多数情况下都与获胜模型一致，因此表现出色。标准 SVM 具有与平衡预测器相似的性能，随着攻击变得激烈，表现不佳。什么时候 $p=0.5$ ，也就是说，当合法电子邮件的一半时间被修改而所有垃圾邮件都被修改时，平衡预测器仍然表现出非常稳定的性能，而平衡预测器 Equi 的性能 ${ }^{* 1}$ 和装备 $^{* 2}$ 在攻击因子增加到 0.3 后立即急剧恶化。混合策略 (同样不是最佳预测器) 在给定任何攻击强度级别的所有预测器中展示了最一致的性能。
我们还测试了攻击因素的情况 $f_{\mathrm{a}} \in(0,1)$ 对于该数据集上的每个受攻击样本，在均匀分布下是完全随机的。负面数据被攻击的概率从 0.1 逐渐增加到 0.9 。结果如图 13.11 所示。同样，我们观察到所有预测变量的相似行为: 稳定均衡预测变量， $\mathrm{SVM}^{* 1}, \mathrm{SVM}^{* 2}$ ，并且 SVM 逐渐恶化为 $p$ 增加。混合策略虽然随着更多负面数据被允许修改而减弱，但始终位于均衡预测变量和 $S V M$ 预测因子。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|ECON2070

Posted on 2023年3月18日2023年3月18日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Equilibrium Strategy Determination

In this game, the defender is considered the controller as it can select the actions which move the game from a state to another. The defender will also control when to apply the MTD, i.e. it will determine the duration between the time steps of the stochastic game. Assuming that the attacker has enough power, it can complete the brute-force attack in time $t_i$ for $i=1,2, \ldots, N$ for each one of the encryption techniques. Then, the defender should choose the time step $t$ to take the next action as follows:
$$
t<\min \left(t_i\right), \quad i=1,2, \ldots, N .
$$
By doing this, the defender can make sure that it takes a timely action before the attacker succeeds in revealing one of the keys.
The accumulated utility of player $i$ at state $s$ will be
$$
\Phi_i(\boldsymbol{f}, \boldsymbol{g}, s)=\sum_{t=1}^{\infty} \beta^{t-1} \cdot U_i\left(f\left(s_t\right), g\left(s_t\right), s_t\right),
$$
where $\boldsymbol{f}$ and $\boldsymbol{g}$ are the strategies adopted by the defender and attacker, respectively. The strategy specifies a vector of actions to be chosen at each of the states, e.g. $\boldsymbol{f}=\left[f\left(s_1\right), \ldots, f\left(s_K\right)\right]$ for all the $K$ states. Actions $f\left(s_t\right)$ and $g\left(s_t\right)$ are the actions chosen at $s_t$, which is the state of the game at time $t$, according to strategies $\boldsymbol{f}, \boldsymbol{g}$. State $s_t \in \mathcal{\delta}$ is determined by the defender’s action at time $t-1$. The game is assumed to start at a specific state $s=s_1$. Note that the utility in (10.4) is always bounded at infinity due to the fact that $0<\beta<1$.

When designing the bimatrix, the defender needs to calculate the accumulated utility when choosing each pure strategy against all of the attacker’s pure strategies. The defender, as a controller, can know the next state resulting from its actions, and, thus, it sums the utilities in all states using the discount factor $\boldsymbol{\beta}$. Let $\boldsymbol{X}$ be the defender’s accumulated utility matrix for all defender’s pure strategies’ permutations and all attacker’s pure strategies’ permutations. We let $\boldsymbol{F}{\boldsymbol{\bullet} \bullet}=\left[\boldsymbol{f}_1, \boldsymbol{f}_2, \ldots, \boldsymbol{f}{K^k}\right]$ be a matrix of all defender’s pure strategies’ permutation where each row represents actions in this strategy and similarly $\boldsymbol{G}{i \bullet}=\left[\boldsymbol{g}_1, \boldsymbol{g}_2, \ldots, \boldsymbol{g}{N^k}\right]$ the matrix of all attacker’s pure strategies’ permutation.

经济代写|博弈论代写Game Theory代考|Simulation Results and Analysis

For our simulations, we choose a system that uses two encryption techniques with two different keys per technique. Thus, the number of system states is four and the defender has four actions in each state. For the bimatrix, the attacker has $2^4=16$ different strategy permutations and the defender has $4^4=256$ different strategy permutations. The power values are set to 1 and 3 to pertain to the ratio between the power consumption in the two different encryption techniques. These values are the same for both players. We set $R_1$ and $R_2$ to be 10 and 5 depending on the opponent’s actions. We choose these values to be higher than the power values in order for the utilities to be positive. The transition reward is set to 5 and 10 for switching to another state defined by another key or another technique, respectively.

First, we run simulations when there is no transition cost, $q=0$. The equilibrium strategies for both the attacker and defender are shown in Table 10.2. Note that actions $a_1, a_2$ represent the selection of two keys for the same encryption technique and actions $a_3, a_4$ represent two keys for another technique. Table 10.1 shows the probabilities over all actions for each player. These probabilities show how players should select actions in every state. For the defender, if it starts in state $s_3$, then it should move to state $s_1$ with the highest probability and move to state $s_2$ with a very similar probability. This is because the defender will change the technique and so gets a higher transition reward. We can see that the probability of moving to the same state is always very low and can reach 0 as in state $s_1$. The probability of moving to a state that has a similar encryption key is always less than that of moving to a state with different technique as the transition reward will be lower. For the attacker, the probability of attacking the same technique that is used in the current state is always higher than attacking any other technique.

In Figure 10.3, we show the effect of the discount factor on the defender’s utility at equilibrium in every state. First, we can see that all utility values at all states increase as the discount factor increases. This is due to the fact that increasing the discount factor will make the defender care more about future rewards thus choosing the actions that will increase these future rewards. Figure 10.3 also shows that the defender’s values at states 1 and 2 are higher than at states 3 and 4 . This because states 1 and 2 adopt the first encryption technique which uses less power than the encryption technique used in states 3 and 4 . The difference mainly arises in the first state before switching to other states and applying the discount factor. Clearly, changing the discount factor has a big effect on changing the equilibrium strategy, and, thus, the game will move between states with different probabilities resulting in a different accumulated reward.

博弈论代考

经济代写|博弈论代写Game Theory代考|Equilibrium Strategy Determination

在此游戏中，防御者被视为控制器，因为它可以选择将游戏从一种状态移动到另一种状态的动作。防御者还将控制何时应用 MTD，即它将确定随机游戏的时间步长之间的持续时间。假设攻击者拥有足够的力量，可以及时完成暴力破解 $t_i$ 为了 $i=1,2, \ldots, N$ 对于每一种加密技术。然后，防御者应该选择时间步长 $t$ 采取下一步行动如下:
$$
t<\min \left(t_i\right), \quad i=1,2, \ldots, N
$$
通过这样做，防御者可以确保在攻击者成功泄露其中一个密钥之前采取及时的行动。玩家的侽积效用 $i$ 在状态 $s$ 将
$$
\Phi_i(\boldsymbol{f}, \boldsymbol{g}, s)=\sum_{t=1}^{\infty} \beta^{t-1} \cdot U_i\left(f\left(s_t\right), g\left(s_t\right), s_t\right)
$$
在哪里 $\boldsymbol{f}$ 和 $\boldsymbol{g}$ 分别是防御者和攻击者所采用的策略。该策略指定了在每个状态下要选择的动作向量，例如 $\boldsymbol{f}=\left[f\left(s_1\right), \ldots, f\left(s_K\right)\right]$ 对于所有的 $K$ 状态。动作 $f\left(s_t\right)$ 和 $g\left(s_t\right)$ 是在 $s_t$ ，这是当时的游戏状态 $t$, 根据策略 $\boldsymbol{f}, \boldsymbol{g}$. 状态 $s_t \in \delta$ 由防御者当时的动作决定 $t-1$. 假定游戏从特定状态开始 $s=s_1$. 请注意，由于以下事实，(10.4) 中的效用总是有界于无穷大 $0<\beta<1$.
在设计双矩阵时，防御者需要计算选择每个纯策略时对攻击者所有纯策略的侽积效用。作为控制器的防御者可以知道其行为导致的下一个状态，因此它使用折扣因子对所有状态下的效用求和 $\beta$. 让 $\boldsymbol{X}$ 是防御者所有防御者纯策略排列和所有攻击者纯策略排列的侽积效用矩阵。我们让 $\boldsymbol{F} \bullet \bullet=\left[\boldsymbol{f}_1, \boldsymbol{f}_2, \ldots, \boldsymbol{f} K^k\right]$ 是所有防御者纯策略排列的矩阵，其中每一行代表该策略中的动作，类似地 $\boldsymbol{G} i \bullet=\left[\boldsymbol{g}_1, \boldsymbol{g}_2, \ldots, \boldsymbol{g} N^k\right]$ 所有攻击者纯策略排列的矩阵。

经济代写|博弈论代写Game Theory代考|Simulation Results and Analysis

对于我们的模拟，我们选择一个系统，该系统使用两种加密技术，每种技术有两个不同的密钥。因此，系统状态的数量为四个，防御者在每个状态下有四个动作。对于双矩阵，攻击者有 $2^4=16$ 不同的策略排列，防御者有 $4^4=256$ 不同的策略排列。功率值设置为 1 和 3 ，以适应两种不同加密技术的功耗比率。这些值对于两个玩家都是相同的。我们设置 $R_1$ 和 $R_2 10$ 和 5 取决于对手的行动。我们选择这些值高于功率值以使效用为正。转换奖励分别设置为 5 和 10 ，用于切换到由另一个键或另一种技术定义的另一个状态。
首先，我们在没有转换成本的情况下运行模拟， $q=0$. 表 10.2 显示了攻击者和防御者的均衡策略。注意动作 $a_1, a_2$ 表示为相同的加密技术和操作选择两个密钥 $a_3, a_4$ 代表另一种技术的两个键。表 10.1 显示了每个参与者所有行动的概率。这些概率显示了玩家在每个状态下应该如何选择动作。对于防御者，如果它开始于状态 $s_3$ ，那么它应该移动到状态 $s_1$ 以最高的概率移动到状态 $s_2$ 概率非常相似。这是因为防守者会改变技术并因此获得更高的过渡奖励。我们可以看到移动到相同状态的概率总是很低，可以达到 $0 s_1$. 移动到具有相似加密密钥的状态的概率总是小于移动到具有不同技术的状态的概率，因为转换奖励会更低。对于攻击者来说，攻击当前状态下使用的相同技术的概率总是高于攻击任何其他技术的概率。
在图 10.3 中，我们显示了贴现因子对防御者在每个状态下的均衡效用的影响。首先，我们可以看到所有状态下的所有效用值都随着贴现因子的增加而增加。这是因为增加折扣因子会使防御者更关心末来的回报，从而选择会增加这些末来回报的行动。图 10.3 还显示防御者在状态 1 和 2 的值高于状态 3 和 4。这是因为状态 1 和 2 采用第一种加密技术，该技术比状态 3 和 4 中使用的加密技术使用更少的功率。差异主要出现在切换到其他状态并应用折扣因子之前的第一个状态。显然，改变贴现因子对改变均衡策略有很大影响。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|ECON3503

Posted on 2023年3月18日2023年3月18日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Numerical Example

Consider a game that has two states. In an MTD scenario, the defender will have two actions representing selecting between each of these two states. Similarly, the attacker is assumed to have two actions of attacking profiles (both can be applied at either state).

A defender’s strategy can then be given as $\boldsymbol{f}=[a b]$, where $a$ is the action of moving to state $s_1$ and $b$ is the action of moving to state $s_2$.
The defender’s and attacker’s strategies’ permutation will be given as:
$$
\boldsymbol{F}=\left[\begin{array}{ll}
1 & 1 \
1 & 2 \
2 & 1 \
2 & 2
\end{array}\right] \boldsymbol{G}=\left[\begin{array}{ll}
1 & 1 \
1 & 2 \
2 & 1 \
2 & 2
\end{array}\right]
$$
where $\boldsymbol{F}$ and $\boldsymbol{G}$ represent the defender’s and the attacker’s permutations, respectively.
Since each player has four different permutations, each of the formulated matrices (bimatrix) will be of size $4 \cdot 4$ representing all the possible combinations of the players’ permutations. The elements of these matrices will be the accumulated utilities for each player resulting from starting at combination and considering all the future transitions.

Now suppose that the mixed Nash equilibrium for the bimatrix game, calculated from any numerical algorithm such as Lemke and Howson (1964), is given by:
$$
\boldsymbol{x}^=\left[\begin{array}{l} 0.2 \ 0.1 \ 0.3 \ 0.4 \end{array}\right] \boldsymbol{y}^=\left[\begin{array}{l}
0.0 \
0.1 \
0.4 \
0.5
\end{array}\right]
$$

where each row in $\boldsymbol{x}^$ and $\boldsymbol{y}^$ represents the players’ probabilities of selecting a strategy in $\boldsymbol{F}$ and $\boldsymbol{G}$, respectively.

Finally, the stochastic game equilibrium strategies can be calculated by summing the probabilities of choosing each action over the corresponding states. For example, the defender can choose action 1 at state 1 twice in $\boldsymbol{F}$ with probabilities 0.2 and 0.1 (from $\boldsymbol{x}^$ ). When summed, the defender knows that when the game is at state 1 , it will choose action 1 with a probability 0.3 . Following the same approach, we can compute the full equilibrium matrices as follows: $$ \boldsymbol{E}^=\left[\begin{array}{ll}
0.3 & 0.5 \
0.7 & 0.5
\end{array}\right] \boldsymbol{H}^=\left[\begin{array}{ll} 0.1 & 0.4 \ 0.9 & 0.6 \end{array}\right] $$ where $\boldsymbol{E}^$ and $\boldsymbol{H}^*$ are the defender’s and the attacker’s equilibrium solutions, respectively, and that the rows of the matrices represent players’ actions and the columns represent the game states.

经济代写|博弈论代写Game Theory代考|A Case Study for Applying Single-Controller

Consider a wireless sensor network that consists of a BS and a number of wireless nodes. The network is deployed for sensing and collecting data about some phenomena in a given geographic area. Sensors will collect data and use multi-hop transmissions to forward this data to a central receiver or BS. The multiple access follows a slotted Aloha protocol. Time is divided into slots and the time slot size equals the time required to process and send one packet. Sensor nodes are synchronized with respect to time slots. We assume that nodes are continuously working and so every time slot there will be data that must be sent to the BS.

All packets sent over the network are assumed to be decrypted using a given encryption technique and a previously shared secret key. All the nodes in the system are pre-programmed with a number of encryption techniques along with a number of encryption keys per technique, as what is typically done in sensor networks (Casola et al. 2013). The BS chooses a specific encryption technique and key by sending a specific control signal over the network including the combination it wants to use. We note that the encryption technique and key sizes should be carefully selected in order not to consume a significant amount of energy when encrypting or decrypting packets. Increasing the key size will increase the amount of consumed energy particularly during the decryption (Lee et al. 2010). Since the BS is mostly receiving data, it spends more time decrypting packets rather than encrypting them and, thus, it will be highly affected by key size selection.

In our model, an eavesdropper is located in the communication field of the BS and it can listen to packets sent or received by the BS. As packets are encrypted, the attacker will seek to decrypt the packets it receives in order to get information. The attacker knows the encryption techniques used in the network and so it can try every possible key on the received packets until getting useful information. This technique is known as brute-force attack.

The idea of using multiple encryption techniques was introduced in Casola et al. (2013). However, in this work, each node individually selects one of these technique to encrypt transmitted packets. The receiving node can know the used technique by a specific field in the packet header. Large encryption keys were used which require a significant amount of power to be decrypted. Nonetheless, these large keys are highly unlikely to be revealed using a brute-force attack in a reasonable time. Here, we propose to use small encryption keys to save energy and, in conjunction with that, we enable the BS to change the encryption method in a way that reduces the chance that the encryption key is revealed by the attacker. This is the main idea behind MTD. In this model, the encryption key represents the attack surface, and by changing the encryption method, the BS will make it harder for the eavesdropper to reveal the key and get the information from the system. Naturally, the goals of the eavesdropper and the BS are not aligned. On the one hand, the BS wants to protect the data sent over the network by changing encryption method. On the other hand, the attacker wants to reveal the used key in order to get information. To understand the interactions between the defender and the attacker, one can use game theory to study their behavior in this MTD scenario. The problem is modeled as a game in which the attacker and the defender are the players. As the encryption method should be changed over time and depending on the attacker’s actions, we must use a dynamic game.

博弈论代考

经济代写|博弈论代写Game Theory代考|Numerical Example

考虑一个有两个状态的游戏。在 MTD 场景中，防御者将有两个动作代表在这两个状态中的每一个之间进行选择。类似地，假设攻击者有两个攻击配置文件的操作（都可以在任一状态下应用）。
防御者的策略可以表示为 $\boldsymbol{f}=[a b]$ ，在哪里 $a$ 是移动到状态的动作 $s_1$ 和 $b$ 是移动到状态的动作 $s_2$. 防御者和攻击者的策略排列如下:
在哪里 $\boldsymbol{F}$ 和 $\boldsymbol{G}$ 分别代表防御者和攻击者的排列。
由于每个玩家都有四种不同的排列，因此每个公式化的矩阵 (双矩阵) 的大小都是 $4 \cdot 4$ 代表玩家排列的所有可能组合。这些矩阵的元素将是每个玩家从组合开始并考虑所有末来转换的侽积效用。
现在假设双矩阵博栾的混合纳什均衡，由 Lemke 和 Howson (1964) 等任何数值算法计算得出，由下式给出:
$$
\boldsymbol{x}^{=}\left[\begin{array}{llll}
0.2 & 0.1 & 0.3 & 0.4
\end{array}\right] \boldsymbol{y}=\left[\begin{array}{llll}
0.0 & 0.1 & 0.4 & 0.5
\end{array}\right]
$$
最后，可以通过对相应状态下选择每个动作的概率求和来计算随机博孪均衡策略。例如，防御者可以在状态 1 中选择动作 1 两次 $\boldsymbol{F}$ 概率为 0.2 和 0.1 (来自 lboldsymbol $x}^{\wedge}$ ). 求和时，防御者知道当游戏处于状态 1 时，它会以 0.3 的概率选择动作 1 。按照相同的方法，我们可以按如下方式计算完整的平衡矩阵:
在哪里 Iboldsymbol{E $}^{\wedge}$ 和 $\boldsymbol{H}^*$ 分别是防御者和攻击者的均衡解，矩阵的行代表玩家的行为，列代表游戏状态。

经济代写|博弈论代写Game Theory代考|A Case Study for Applying Single-Controller

考虑一个由 BS 和许多无线节点组成的无线传感器网络。该网络用于感测和收集有关给定地理区域中某些现象的数据。传感器将收集数据并使用多跳传输将此数据转发到中央接收器或 BS。多路访问遵循时隙 Aloha 协议。时间分为时隙，时隙大小等于处理和发送一个数据包所需的时间。传感器节点在时隙方面同步。我们假设节点连续工作，因此每个时隙都会有必须发送到 BS 的数据。

假设通过网络发送的所有数据包都使用给定的加密技术和先前共享的密钥进行解密。系统中的所有节点都使用多种加密技术以及每种技术的多种加密密钥进行预编程，这与传感器网络中的典型做法相同（Casola 等人，2013 年）。BS 通过在网络上发送特定的控制信号（包括它要使用的组合）来选择特定的加密技术和密钥。我们注意到，应仔细选择加密技术和密钥大小，以免在加密或解密数据包时消耗大量能量。增加密钥大小会增加消耗的能量，尤其是在解密期间（Lee 等人，2010 年）。由于 BS 主要接收数据，

在我们的模型中，窃听器位于 BS 的通信领域，它可以监听 BS 发送或接收的数据包。由于数据包被加密，攻击者将试图解密它收到的数据包以获取信息。攻击者知道网络中使用的加密技术，因此它可以在收到的数据包上尝试每个可能的密钥，直到获得有用的信息。这种技术被称为蛮力攻击。

Casola 等人介绍了使用多种加密技术的想法。(2013)。然而，在这项工作中，每个节点单独选择其中一种技术来加密传输的数据包。接收节点可以通过包头中的特定字段知道所使用的技术。使用了需要大量能量才能解密的大型加密密钥。尽管如此，这些大密钥极不可能在合理的时间内使用暴力攻击来泄露。在这里，我们建议使用小的加密密钥来节省能源，与此同时，我们使 BS 能够以一种降低加密密钥被攻击者泄露的机会的方式改变加密方法。这是 MTD 背后的主要思想。在这个模型中，加密密钥代表攻击面，通过改变加密方式，BS 将使窃听者更难泄露密钥并从系统中获取信息。自然地，窃听者和 BS 的目标并不一致。一方面，BS 想通过改变加密方法来保护通过网络发送的数据。另一方面，攻击者想要泄露使用的密钥以获取信息。要了解防御者和攻击者之间的交互，可以使用博弈论来研究他们在此 MTD 场景中的行为。该问题被建模为一个游戏，其中攻击者和防御者是玩家。由于加密方法应该随着时间的推移而改变，并且取决于攻击者的行为，我们必须使用动态游戏。窃听者和 BS 的目标不一致。一方面，BS 想通过改变加密方法来保护通过网络发送的数据。另一方面，攻击者想要泄露使用的密钥以获取信息。要了解防御者和攻击者之间的交互，可以使用博弈论来研究他们在此 MTD 场景中的行为。该问题被建模为一个游戏，其中攻击者和防御者是玩家。由于加密方法应该随着时间的推移而改变，并且取决于攻击者的行为，我们必须使用动态游戏。窃听者和 BS 的目标不一致。一方面，BS 想通过改变加密方法来保护通过网络发送的数据。另一方面，攻击者想要泄露使用的密钥以获取信息。要了解防御者和攻击者之间的交互，可以使用博弈论来研究他们在此 MTD 场景中的行为。该问题被建模为一个游戏，其中攻击者和防御者是玩家。由于加密方法应该随着时间的推移而改变，并且取决于攻击者的行为，我们必须使用动态游戏。要了解防御者和攻击者之间的交互，可以使用博弈论来研究他们在此 MTD 场景中的行为。该问题被建模为一个游戏，其中攻击者和防御者是玩家。由于加密方法应该随着时间的推移而改变，并且取决于攻击者的行为，我们必须使用动态游戏。要了解防御者和攻击者之间的交互，可以使用博弈论来研究他们在此 MTD 场景中的行为。该问题被建模为一个游戏，其中攻击者和防御者是玩家。由于加密方法应该随着时间的推移而改变，并且取决于攻击者的行为，我们必须使用动态游戏。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

金融工程代写

非参数统计代写

非参数统计指的是一种统计方法，其中不假设数据来自于由少数参数决定的规定模型；这种模型的例子包括正态分布模型和线性回归模型。

广义线性模型代考

广义线性模型（GLM）归属统计学领域，是一种应用灵活的线性回归模型。该模型允许因变量的偏差分布有除了正态分布之外的其它分布。

有限元方法代写

随机分析代写

时间序列分析代写

回归分析代写

MATLAB代写

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写

经济代写|博弈论代写Game Theory代考|ECON3503

Posted on 2023年3月15日2023年3月15日 by statistics-lab

如果你也在怎样代写博弈论Game Theory这个学科遇到相关的难题，请随时右上角联系我们的24/7代写客服。

我们提供的博弈论Game Theory及其相关学科的代写，服务范围广, 其中包括但不限于:

Statistical Inference 统计推断
Statistical Computing 统计计算
Advanced Probability Theory 高等概率论
Advanced Mathematical Statistics 高等数理统计学
(Generalized) Linear Models 广义线性模型
Statistical Machine Learning 统计机器学习
Longitudinal Data Analysis 纵向数据分析
Foundations of Data Science 数据科学基础

经济代写|博弈论代写Game Theory代考|Hypergames on Graphs

A hypergame models the situation where different players perceive their interaction with other players differently and consequently play different games in their own minds depending on their perception. We consider the case where the difference in perception arises because of incomplete and potentially incorrect information. For instance, suppose a subset of nodes in the network are honeypots, the attacker may mistake these to be true hosts. We formulate a hypergame to model the interaction between the defender and the attacker given asymmetric information.
First, let’s review the definition of hypergames.
Definition 6.2 (Hypergame; Bennett 1980; Vane 2000) Given two players, a game perceived by player 1 is denoted by $\mathcal{G}_1$, and a game perceived by player 2 is denoted by $\mathcal{G}_2$. A level-1 hypergame is defined as a tuple:
$$
\mathcal{H} \mathcal{G}^1=\left\langle\mathcal{G}_1, \mathcal{G}_2\right\rangle
$$
In a level-1 hypergame, none of the player’s is aware of other player’s perception.
When one player becomes aware of the other player’s (mis)perception, the interaction is captured by a level- 2 two-player hypergame, defined as a tuple:
$$
\mathcal{H} \mathcal{G}^2=\left\langle\mathcal{H} \mathcal{G}^1, \mathcal{G}_2\right\rangle
$$
where P1 perceives the interaction as a level-1 hypergame (as P1 is aware of P2’s game $\mathcal{G}_2$ in addition to his own) and $\mathrm{P} 2$ perceives the interaction as the game $\mathcal{G}_2$.

We refer to the games $\mathcal{G}_1$ (resp., $\mathcal{G}_2$ ) as P1’s (resp., P2’s) perceptual game in level-1 hypergame, and $\mathcal{H} \mathcal{G}^1$ as P1’s perceptual game in level-2 hypergame. As P2 is not aware that she might be misperceiving the game, her perceptual game in level-2 hypergame is still $\mathcal{G}_2$.

In general, if P1 computes his strategy by solving an $(m-1)$ th level hypergame and P2 computes her strategy using an $n$th level hypergame with $n<m$, then the resulting hypergame is said to be a level- $m$ hypergame given as:
$$
\mathcal{H} \mathcal{G}^m=\left\langle\mathcal{H} \mathcal{G}_1^{m-1}, \mathcal{H} \mathcal{G}_2^n\right\rangle
$$

经济代写|博弈论代写Game Theory代考|Synthesis of Reactive Defense Strategies with Cyber Deception

When the attacker has a one-sided misperception of labeling function, as defined in Definition 6.3, the defender might strategically utilize this misperception to deceive the attacker into choosing a strategy that is advantageous to the defender. To understand when the defender might have such a deceptive strategy and how to compute it, we study the solution concept of hypergame.

Solution Approach A hypergame $\mathcal{H} \mathcal{G}^2=\left\langle\mathcal{H} \mathcal{G}^1, \mathcal{G}_2\right\rangle$ is defined using two games, namely $\mathcal{G}_1$ and $\mathcal{G}_2$. Under one-sided misperception of labeling function, defender is aware of both games. Therefore, to synthesize a deceptive strategy, the defender must take into account the strategy that the attacker will use, based on her misperception. That is, the defender must solve two games: game $\mathcal{G}_2$ to identify the set of states in the game transition system $G_2 \otimes \mathcal{A}$ that the attacker perceives as winning for her under labeling function $L_2$ and game $\mathcal{G}_1$ to identify the set of states in the game transition system $G_1 \otimes \mathcal{A}$ that are winning for the defender under (ground-truth) labeling function $L=L_1$. After solving the two games, the defender can integrate the solutions to obtain a set of states, at which the attacker makes mistakes due to the difference between $L_2$ and $L$. Let us introduce a notation to denote these sets of winning states.

$\mathcal{G}_1$ : P1’s winning region is Win $\operatorname{Win}_1 \subseteq S \times Q$ and P2’s winning region is Win $_2 \subseteq S \times Q$.
$\mathcal{G}_2: \mathrm{P} 1$ ‘s winning region is $\mathrm{Win}_1^{\mathrm{P2}} \subseteq S \times Q$ and P2’s winning region is $\mathrm{Win}_2^{\mathrm{P2}} \subseteq S \times Q$.
Figure 6.7 provides a conceptual representation partitions of the state space of a game transition system. Due to misperception, the set of states are partitioned into the following regions:
Win $\mathrm{W}_1$ : is a set of states from which P1 can ensure satisfaction of security objectives, even if P2 has complete and correct information. Thus, P1 can take the winning strategy $\pi_1$.
$\mathrm{Win}_1^{\mathrm{P} 2} \cap \mathrm{Win}_2$ is a set of states where P2 is truly winning but perceives the states to be losing for her due to misperception. Thus, P2 may either give up the attack mission or play randomly.
$\mathrm{Win}_2^{\mathrm{P2}} \cap \mathrm{Win}_2$ is a set of states in which P2 is truly winning and perceives those states to be winning. In this scenario, she will carry out the winning strategy $\pi_2^{\mathrm{P2}}$. However, this strategy can be different from the true winning strategy $\pi_2$ that P2 should have played if she had complete and correct information. This difference creates unique opportunities for P1 to enforce security of the system.

博弈论代考

经济代写|博弈论代写Game Theory代考|Hypergames on Graphs

超级博亦模拟了不同玩家对他们与其他玩家的互动有不同看法的情况，因此根据他们的看法在他们自己的头脑中玩不同的游戏。我们考虑由于不完整和可能不正确的信息而导致感知差异的情况。例如，假设网络中的一部分节点是蜜罐，攻击者可能会将它们误认为是真正的主机。我们制定了一个超博娈来模拟给定信息不对称的防御者和攻击者之间的交互。
首先，让我们回顾一下超博栾的定义。
定义 6.2 (Hypergame；Bennett 1980；Vane 2000) 给定两个玩家，玩家 1 感知的游戏表示为 $\mathcal{G}_1$ ，玩家 2 感知到的游戏表示为 $\mathcal{G}_2$.一级超博变被定义为一个元组:
$$
\mathcal{H \mathcal { G } ^ { 1 }}=\left\langle\mathcal{G}_1, \mathcal{G}_2\right\rangle
$$
在 1 级超博亦中，没有玩家知道其他玩家的感知。
当一个玩家意识到另一个玩家的（错误）感知时，交互被 2 级双人超博恋捕获，定义为元组:
$$
\mathcal{H} \mathcal{G}^2=\left\langle\mathcal{H} \mathcal{G}^1, \mathcal{G}_2\right\rangle
$$
其中 $\mathrm{P} 1$ 将交互视为 1 级超博恋 (因为 $\mathrm{P} 1$ 知道 $\mathrm{P} 2$ 的游戏 $\mathcal{G}_2$ 除了他自己的) 和 $\mathrm{P} 2$ 将交互视为游戏 $\mathcal{G}_2$.
我们指的是游戏 $\mathcal{G}_1$ (分别， $\mathcal{G}_2$ ) 作为 P1 (resp., P2) 在 level-1 hypergame 中的感知游戏，并且 $\mathcal{H} \mathcal{G}^1$ 作为 2 级超博娈中 $P 1$ 的感知博恋。由于 $P 2$ 不知道她可能误解了游戏，她在 level-2 hypergame 中的感知游戏仍然是 $\mathcal{G}_2$.
一般来说，如果 P1 通过解决一个问题来计算他的策略 $(m-1)$ th level hypergame 和 P2 计算她的策略使用 $n$ th级超博恋 $n<m$ ，那么由此产生的超博恋被称为水平- $m$ 超博栾给出：
$$
\mathcal{H} \mathcal{G}^m=\left\langle\mathcal{H} \mathcal{G}_1^{m-1}, \mathcal{H} \mathcal{G}_2^n\right\rangle
$$

经济代写|博弈论代写Game Theory代考|Synthesis of Reactive Defense Strategies with Cyber Deception

当攻击者对标签函数有片面的误解时，如定义 6.3 所定义，防御者可能会战略性地利用这种误解来欺骗攻击者选择对防御者有利的策略。为了了解防御者何时可能采用这种欺骗性策略以及如何计算它，我们研究了超博弈的解决方案概念。
解决方法超博栾 $\mathcal{H} \mathcal{G}^2=\left\langle\mathcal{H} \mathcal{G}^1, \mathcal{G}_2\right\rangle$ 使用两个游戏定义，即 $\mathcal{G}_1$ 和 $\mathcal{G}_2$. 在对标签功能的片面误解下，防守者意识到了这两种游戏。因此，要综合欺骗性策略，防御者必须考虑攻击者将根据自己的误解使用的策略。也就是说，防守者必须解决两场比寋: 比赛 $\mathcal{G}_2$ 识别游戏转换系统中的状态集 $G_2 \otimes \mathcal{A}$ 攻击者认为在标签功能下为她赢得胜利 $L_2$ 和游戏 $\mathcal{G}_1$ 识别游戏转换系统中的状态集 $G_1 \otimes \mathcal{A}$ 在 (真实情况) 标签功能下为防御者赢得胜利 $L=L_1$. 解决完这两个游戏后，防御者可以将解决方案整合以获得一组状态，攻击者由于之间的差异而犯错误 $L_2$ 和 $L$. 让我们引入一个符号来表示这些获胜状态集。

$\mathcal{G}_1: \mathrm{P} 1$ 的获胜区域是WinWin $\operatorname{Win}_1 \subseteq S \times Q \mathrm{P} 2$ 的获胜区域是 $\operatorname{Win}_2 \subseteq S \times Q$.
$\mathcal{G}_2: \mathrm{P} 1$ 的获胜区域是 $\mathrm{Win}_1^{\mathrm{P} 2} \subseteq S \times Q \mathrm{P} 2$ 的获胜区域是 $\mathrm{Win}_2^{\mathrm{P} 2} \subseteq S \times Q$. 图 6.7 提供了游戏转换系统状态空间的概念表示分区。由于误解，状态集被划分为以下区域:
赢 $W_1$ : 是一组状态，即使 $P 2$ 拥有完整和正确的信息，P1 也可以确保满足安全目标。因此，P1 可以釆取必胜策略 $\pi_1$.
$\operatorname{Win}_1^{\mathrm{P2}} \cap \mathrm{Win}$ 是一组状态，其中 $P 2$ 真正获胜，但由于误解而认为这些状态对她来说是失败的。因此，P2 要么放弃攻击任务，要么随机游玩。
$\mathrm{Win}_2^{\mathrm{P} 2} \cap \mathrm{Win}_2$ 是 $\mathrm{P} 2$ 真正获胜并认为这些状态获胜的一组状态。在这种情况下，她将执行制胜战略 $\pi_2^{\mathrm{P} 2}$. 但是，此策略可能与真正的获胜策略不同 $\pi_2$ 如果她有完整和正确的信息， $P 2$ 应该玩。这种差异为 P1 创造了独特的机会来加强系统的安全性。

统计代写请认准statistics-lab™. statistics-lab™为您的留学生涯保驾护航。

R语言代写	问卷设计与分析代写
PYTHON代写	回归分析与线性模型代写
MATLAB代写	方差分析与试验设计代写
STATA代写	机器学习/统计学习代写
SPSS代写	计量经济学代写
EVIEWS代写	时间序列分析代写
EXCEL代写	深度学习代写
SQL代写	各种数据建模与可视化代写