统计代写|强化学习作业代写Reinforcement Learning代考|Reinforcement learning

statistics-lab™ 为您的留学生涯保驾护航 在代写强化学习Reinforcement Learning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。我们的专家在代写强化学习Reinforcement Learning代写方面经验极为丰富，各种代写强化学习Reinforcement Learning相关的作业也就用不着说。

• Statistical Inference 统计推断
• Statistical Computing 统计计算
• (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

统计代写|强化学习作业代写Reinforcement Learning代考|Reinforcement learning

Reinforcement learning can be used to solve very big problems with many discrete state configurations or problems with continuous state space. Consider the game of backgammon, which has close to $10^{20}$ discrete states, or consider the game of Go, which has close to $10^{170}$ discrete states. Also consider environments like self-driving cars, drones, or robots: these have a continuous state space.
Up to now we saw problems where the state space was discrete and also small in size, such as the grid world with $\sim 100$ states or the taxi world with 500 states. How do we scale the algorithms we have learned so far to bigger environments or environments with continuous state spaces? All along we have been representing the state values $V(s)$ or the action values $Q(s, a)$ with a table, with one entry for each value of state $s$ or a combination of state $s$ and action $a$. As the numbers increase, the table size is going to become huge, making it infeasible to be able to store state or action values in a table. Further, there will be too many combinations, which can slow down the learning of a policy. The algorithm may spend too much time in states that are very low probability in a real run of the environment.

We will take a different approach now. Let’s represent the state value (or state-action value) with the following function:
$$\begin{gathered} \hat{v}(s ; w) \approx v_{\pi}(s) \ \hat{q}(s, a ; w) \approx q_{\pi}(s, a) \end{gathered}$$
Instead of representing values in a table, they are now being represented by the function $\hat{v}(s ; w)$ or $\hat{q}(s, a ; w)$ where the parameter $w$ is dependent on the policy being followed by the agent, and where $s$ or $(s, a)$ are the inputs to the state or state-value functions. We choose the number of parameters $|w|$ which is lot smaller than the number of states $|s|$ or the number of state-action pairs $(|s| x|a|)$. The consequence of this approach is that there is a generalization of representation of state of the stateaction values. When we update the weight vector $w$ based on some update equation for a given state $s$, it not only updates the value for that specific $s$ or $(s, a)$, but also updates the values for many other states or state actions that are close to the original $s$ or $(s, a)$ for which the update has been carried out. This depends on the geometry of the function. The other values of states near $s$ will also be impacted by such an update as shown previously. We are approximating the values with a function that is a lot more restricted than the number of states. Just to be specific, instead of updating $v(s)$ or $q(s, a)$ directly, we now update the parameter set $w$ of the function, which in turn impacts the value estimates $\hat{v}(s ; w)$ or $\hat{q}(s, a ; w)$. Of course, like before, we carry out the $w$ update using the MC or TD approach. There are various approaches to function approximation. We could feed the state vector (the values of all the variables that signify the state, e.g., position, speed, location, etc.) and get $\hat{v}(s ; w)$, or we could feed state and action vectors and get $\hat{q}(s, a ; w)$ as an output. An alternate approach that is very dominant in the case of actions being discrete and coming from a small set is to feed state vector $s$ and get $|A|$ number of $\hat{q}(s, a ; w)$, one for each action possible $(|A|$ denotes the number of possible actions). Figure 5-1 shows the schematic.

统计代写|强化学习作业代写Reinforcement Learning代考|Theory of Approximation

Function approximation is a topic studied extensively in the field of supervised learning wherein based on training data we build a generalization of the underlying model. Most of the theory from supervised learning can be applied to reinforcement learning with functional approximation. However, RL with functional approximation brings to fore new issues such as how to bootstrap as well as its impact on nonstationarity. In supervised learning, while the algorithm is learning, the problem/model from which the training data was generated does not change. However, when it comes to RL with function approximation, the way the target (labeled output in supervised learning) is formed, it induces nonstationarity, and we need to come up with new ways to handle it. What we mean by nonstationarity is that we do not know the actual target values of $v(s)$ or $q(s, a)$. We use either the MC or TD approach to form estimates and then use these estimates as “targets.” And as we improve our estimates of target values, we used the revised estimates as new targets. In supervised learning it is different; the targets are given and fixed during training. The learning algorithm has no impact on the targets. In reinforcement learning, we do not have actual targets, and we are using estimates of the target values. As these estimates change, the targets being used in the learning algorithm change; i.e., they are not fixed or stationary during the learning.
Let’s revisit the update equations for $\mathrm{MC}$ (equation 4.2) and TD (equation 4.4), reproduced here. We have modified the equations to make both MC and TD use the same notations of subscript $t$ for the current time and $t+1$ for the next instant. Both equations carry out the same update to move $V_{t}(s)$ closer to its target, which is $G_{t}(s)$ in the case of the $\mathrm{MC}$ update and $R_{t+1}+\gamma * V_{t}(s)$ for the $\operatorname{TD}(0)$ update.
$$\begin{gathered} V_{t+1}(s)=V_{t}(s)+\alpha\left[G_{t}(s)-V_{t}(s)\right] \ V_{t+1}(s)=V_{t}(s)+\alpha\left[R_{t+1}+\gamma * V_{t}\left(s^{\prime}\right)-V_{t}(s)\right] \end{gathered}$$
This is similar to what we do in supervised learning, especially in linear least square regression. We have the output values/targets $y(t)$, and we have the input features $x(t)$, together called training data. We can choose a model Model $_{w}[x(t)]$ like the polynomial linear model, decision tree, or support vectors, or even other nonlinear models like neural nets. The training data is used to minimize the error between what the model is predicting and what the actual output values are from the training set. The is called the minimizing loss function and is represented as follows.

统计代写|强化学习作业代写Reinforcement Learning代考|Coarse Coding

Let’s look at the mountain car problem that was discussed in Figure 2-2. The car has a two-dimensional state, a position, and a velocity. Suppose we divide the twodimensional state space into overlapping circles with each circle representing a feature. If state $S$ lies inside a circle, that particular feature is present and has a value of 1 ; otherwise, the feature is absent and has a value of 0 . The number of features is the number of circles. Let’s say we have $p$ circles; then we have converted a two-dimensional continuous state space to a p-dimensional state space where each dimension can be 0 or 1. In other words, each dimension can belong to ${0,1}$.ellipses, the generalization will be more in the direction of the elongation. We could also choose shapes other than circles to control the amount of generalization.

Now consider the case with large, densely packed circles. A large circle makes the initial generalization wide where two faraway states are connected because they fall inside at least one common circle. However, the density (i.e., number of circles) allows us to control the fine-grained generalization. By having many circles, we ensure that even nearby states have at least one feature that is different between two states. This will hold even when each of the individual circles is big. With the help of experiments with varying configurations of the circle size and number of circles, one can fine-tune the size and number of circles to control the generalization appropriate for the problem/domain in question.

有限元方法代写

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。

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。