自适应学习系统

Have you ever started a course, but thought it was too slow? Or too difficult? Wish you could make it go faster? Felt like you didn't get enough practice to master the content? Adaptive learning systems seek to address these challenges.

您是否曾经上过一门课程，但认为它太慢了？ 还是太难了？ 希望您可以使它运行得更快吗？ 觉得您没有足够的练习来掌握内容吗？ 自适应学习系统试图解决这些挑战。

In this article, I'll go over what adaptive learning systems are. I cover some background on why adaptive learning systems have the structure they do. I'll also introduce a few adaptive learning systems. Then, I'll talk about the four elements of an adaptive learning system, and how you can architect one yourself. We'll wrap up with evaluating the pros and cons of adaptive learning.

在本文中，我将介绍什么是自适应学习系统。 我将介绍一些有关自适应学习系统为何具有其结构的背景知识。 我还将介绍一些自适应学习系统。 然后，我将讨论自适应学习系统的四个要素，以及如何自己构建一个。 我们将总结一下自适应学习的利弊。

什么是适应性学习系统？ (What is an adaptive learning system?)

An adaptive learning system is software where algorithms optimize the content to adjust for the learner's goals and current state of knowledge.

自适应学习系统是一种软件，其中的算法会优化内容以适应学习者的目标和知识的当前状态。

In a traditional e-learning course, you will linearly follow the path an instructor creates. You watch videos, read articles, take quizzes, and practice interactive modules in a predetermined ordered. An adaptive learning system will contain the same types of materials. But the order will change for each learner. The system decides which content to show the learner based on two things:

在传统的电子学习课程中，您将线性遵循教师创建的路径。 您可以按预定顺序观看视频，阅读文章，参加测验并练习互动模块。 自适应学习系统将包含相同类型的材料。 但是顺序会因每个学习者而改变。 系统基于两件事来决定向学习者显示哪些内容：

• If the learner's goal is only a subset of the content, the system can limit the content.

  如果学习者的目标只是内容的一部分，则系统可以限制内容。

• The prior knowledge also comes into play. If the system determines the current path is too easy, the system can speed up to more challenging material. If the system finds out the current path is too difficult, the system may... intervene and review prerequisite content, reduce the challenge, or slow down the pace.

  先验知识也起作用。 如果系统确定当前路径太简单，则系统可以加快处理更具挑战性的材料的速度。 如果系统发现当前路径太困难，则系统可能...干预并检查先决条件内容，减少挑战或减慢步伐。

Some related topics include intelligent tutors, adaptive testing, psychometrics, personalized learning, and smart teaching. Many of these topics share algorithms and structures with adaptive learning systems.

一些相关主题包括智能导师，适应性测试，心理测验，个性化学习和智能教学。 其中许多主题与自适应学习系统共享算法和结构。

知识就是图表：神经科学 (Knowledge is a graph: neuroscience)

I'm going to start with a little background. This will create context for why adaptive learning systems have the four elements below. The point for this is knowledge is a graph.

我将从一些背景开始。 这将为为什么自适应学习系统具有以下四个要素创造背景。 关键是知识就是图表 。

The human brain has 86 billion neurons. Every neuron has dendrites, a soma, and an axon.

人脑有860亿个神经元。 每个神经元都有树突，一个体细胞和一个轴突。

• The dendrites are the input. The edges of the dendrites receive neurotransmitters from the synapse. The synapse is a gap between two neurons.

  树突是输入 。 树突的边缘从突触接收神经递质。 突触是两个神经元之间的间隙。

• The soma is the throughput. The soma -- which contains the cell nucleus -- routes the input from the dendrites.

  躯体就是吞吐量 。 体细胞-包含细胞核-路由来自树突的输入。

• The axon is the output. The axon transmits an action potential -- an electrical signal -- to the axon terminals. A myelin sheath covers the axon to protect the signal. The axon terminals release neurotranmitters into the synapse.

  轴突是输出 。 轴突将动作电位(电信号)传输到轴突端子。 髓鞘覆盖轴突以保护信号。 轴突末端将神经递质释放到突触中。

Information your brain receives and processes corresponds with a neural pathway. Your brain with myelinate that pathway -- strengthen the myelin around the axon to support electrical signals. Because of the strengthened myelin, this path will be more likely to fire in the future. In other words, you learn.

您的大脑接收和处理的信息与神经通路相对应。 您的大脑具有髓鞘通路-增强轴突周围的髓磷脂以支持电信号。 由于髓磷脂的增强，将来这条道路将更有可能开火。 换句话说，您将学习。

Even in the smallest scale, our brain is a massive graph of connected neurons. We learn and optimize by making some paths more likely to connect than other paths.

即使在最小的范围内，我们的大脑也是连接神经元的巨大图形。 我们通过使某些路径比其他路径更可能连接来学习和优化。

知识就是图：学习科学 (Knowledge is a graph: learning science)

The strongest predictor of how we perform in a learning environment is our prior knowledge. What we already know before we start the learning experience. A notable psychology paper -- 1999 Dochy, Segers, and Buehl -- found prior knowledge is 81% of outcome differences between learners. Reviewing prior knowledge before showing new information impacts learning outcomes. And connecting new knowledge to prior knowledge while teaching can have a big impact too. (See Eight Ideas for sources.)

关于我们在学习环境中的表现的最强预测指标是我们的先验知识。 在开始学习之前，我们已经知道了什么。 著名的心理学论文-1999 Dochy，Segers和Buehl-发现先验知识占学习者之间结果差异的81％。 在显示新信息之前回顾先前的知识会影响学习成果。 在教学过程中将新知识与先验知识联系起来也会产生很大的影响。 (有关信息，请参见八个想法 。)

The most famous psychology paper is 1956 "The Magical Number Seven, Plus or Minus Two" by George Miller. The paper suggests that humans have a limited working memory. Miller found for simple numbers, a human could work with about seven items at once. Later researchers found for more complex information, that limit is closer to four.

最著名的心理学论文是乔治·米勒(George Miller)于1956年发表的“魔术数字七，加减二”。 该论文表明，人类的工作记忆有限。 米勒发现，对于简单的数字，一个人可以一次处理大约七个项目。 后来的研究人员发现，对于更复杂的信息，该限制接近于四个。

Some psychologists suggest of these "four slots", for us to learn, at least one or two must be prior knowledge. How much prior knowledge we can "load up" into one of the four slots depends on the strength of the connections in the graph. When we have both prior knowledge and new knowledge in our working memory, we associate the information. And we strengthen the connection between the two. Trying to learn new information without connecting to prior knowledge limits the strength of the memory.

一些心理学家建议这些“四个槽”，供我们学习，至少有一两个必须是先验知识。 我们可以“加载”到四个插槽之一中的多少先验知识取决于图中连接的强度。 当我们在工作记忆中同时拥有先验知识和新知识时，我们便会将这些信息关联起来。 而且我们加强了两者之间的联系。 尝试学习新信息而又不依赖于先验知识会限制存储的强度。

In short, we learn by connecting prior knowledge to new information. And those connections form a large, endless graph of knowledge.

简而言之，我们通过将先验知识与新信息联系起来进行学习。 这些联系形成了无尽的知识大图。

一些重要的自适应学习系统 (A few important adaptive learning systems)

This section is more context, but optional. I'm not writing an thorough article about the history of these systems, but here's some bullets:

本节的内容更多，但可选。 我没有写有关这些系统历史的详尽文章，但是这里有一些要点：

• One of the earliest implementations was the Skinner teaching machine.

  最早的实现之一是Skinner教学机 。

• During the 1960s and 1970s, there were several attempts at computerized instructional systems. Costs and slower machines limited the success of these systems.在1960年代和1970年代，进行了计算机教学系统的几次尝试。 成本和较慢的机器速度限制了这些系统的成功。

• During the late 70s and early 80s, Item Response Theory enabled test makers to start work on computerized adaptive testing.

  在70年代末和80年代初， 项目响应理论使测试人员能够着手进行计算机化自适应测试。

• An early and influential computerized system was the Lisp tutor, also known as LISPITS (1983) at Carnegie Mellon University.卡内基梅隆大学的Lisp导师(也称为LISPITS(1983))是一种早期且有影响力的计算机化系统。

• SuperMemo, released in 1985, incorporated spaced learning into a computerized system.

  1985年发布的SuperMemo将间隔学习功能集成到计算机系统中。

• Also in 1985 came paper for Knowledge Spaces, which forms the foundations of one of the four elements.

  也是在1985年，发表了有关“ 知识空间”的论文，该论文构成了四个要素之一的基础。

• ALEKS Math tutor came out in 1994, heavily promoting its use of knowledge spaces.

  ALEKS数学导师于1994年问世，大力推广其对知识空间的使用。

• In 1995, Corbett and Anderson published "Knowledge tracing", forming the foundation for Bayesian knowledge tracing models.

  1995年，Corbett和Anderson出版了《知识追踪》，为贝叶斯知识追踪模型奠定了基础。

• Some important software includes AutoTutor, ACT-R, and Cognitive Tutor Authoring Tools.

  一些重要的软件包括AutoTutor ， ACT-R和Cognitive Tutor Authoring Tools 。

• Knewton is an example of contemporary adaptive learning systems. Kaplan and Pearson both use Knewton to provide adaptive learning experiences.

  Knewton是当代自适应学习系统的一个例子。 Kaplan和Pearson都使用Knewton提供自适应的学习体验。

四个要素 (The four elements)

Most adaptive learning systems today have these four elements. The terms change and so do their scope. But you will almost always find all four elements.

当今大多数自适应学习系统都具有这四个要素。 术语会发生变化，其范围也会发生变化。 但是您几乎总是会找到所有四个元素。

These elements are:

这些元素是：

• The expert -- a graphical model of the "ideal" state, of everything the person could learn using this system.

  专家 -“理想”状态的图形模型，此人可以使用该系统学习的所有内容。

• The learner -- a model of the learner's current state, which shows how likely the learner is to know each of the nodes in the expert graph.

  学习者 -学习者当前状态的模型，它显示学习者知道专家图中每个节点的可能性。

• The tutor -- the algorithms that determine what content to show and when. The expert model and the learner model inform the tutor. The tutor seeks to optimize content for relevance, challenge, and efficiency.

  导师 -确定何时显示什么内容的算法。 专家模型和学习者模型通知导师。 导师寻求针对相关性，挑战性和效率来优化内容。

• The interface -- which is how to display the learning experience to the learner. In many adaptive learning experience, the interface changes based on the learner model and the tutor's goals.

  界面 -这是如何向学习者显示学习经验的界面 。 在许多自适应学习经验中，界面根据学习者模型和导师的目标而变化。

Let's go into each element.

让我们进入每个元素。

专家-一切的大图 (The expert — the big graph of everything)

The expert model is a large, connected graph of everything you want the learners to know. As the name suggests, you have an expert on the topic -- or experts on topics -- to create the model. This model is static. The expert model only changes when the scope of learning outcomes change. Or when problems and opportunities to refine the adaptive learning system arise.

专家模型是一个大型的，相互连接的图表，其中包含您希望学习者知道的所有内容。 顾名思义，您需要一个主题专家或一个主题专家来创建模型。 此模型是静态的。 专家模型仅在学习结果的范围发生变化时才会发生变化。 或者，当出现问题和机会来完善自适应学习系统时。

Most of the work of the expert model is at the beginning of building a new learning experience. The adaptive learning system will access the expert model to compare the learner's current state with the expert model. The system will also access the expert model to determine which learning experience to focus on next.

专家模型的大部分工作是在建立新的学习体验的开始。 自适应学习系统将访问专家模型，以将学习者的当前状态与专家模型进行比较。 该系统还将访问专家模型，以确定接下来要重点关注的学习经验。

Usually, a team of experts will define the scope of learning outcomes. Each node in the expert model should have the following attributes:

通常，一个专家团队将确定学习成果的范围。 专家模型中的每个节点应具有以下属性：

• A name一个名字
• A short description, which indicates which skills are under test and what is outside the scope简短说明，指出要测试的技能以及范围之外的技能
• A list of prerequisite nodes -- these form the "edges" of the graph. These prerequisites cannot form a "cycle" -- a loop of nodes.必备节点列表-这些构成了图的“边”。 这些先决条件不能形成“循环”-节点循环。

Expert models perform better when each node is small and narrowly defined. For example, each skill in Bloom's taxonomy -- recognition, understanding, application, analysis, synthesis, and evaluation -- could each be its own node in the expert model. The combination of two underlying skills should also be a separate node.

当每个节点较小且定义狭窄时，专家模型的性能会更好。 例如，布卢姆分类法中的每个技能-识别，理解，应用，分析，综合和评估-都可以成为专家模型中自己的节点。 两个基本技能的组合也应该是一个单独的节点。

There is an endless number of formats you could use to create an expert model, such as XML, JSON, CSV, or YAML. It can help to be able to display the expert model graphically for review.

您可以使用多种格式来创建专家模型，例如XML，JSON，CSV或YAML。 能够以图形方式显示专家模型以供审查可能会有所帮助。

Some systems will automatically generate an expert model by querying experts in a series of questions, like a wizard. Others will cluster existing learning content, using algorithms like k-means clustering. You may want to review the Wikipedia article on Knowledge spaces for a more mathematical description.

某些系统会通过查询一系列问题的专家来自动生成专家模型，例如向导。 其他人将使用k-means聚类算法对现有的学习内容进行聚类。 您可能需要查看有关知识空间的Wikipedia文章，以获得更多数学描述。

学习者-您在哪里与您想去哪里 (The learner — where you are versus where you want to be)

The learner element is a model of the learner's current state of ability. So for each given node in the expert graph, the learner model has a probability associated with it: 1-99%. The system updates this graph every time the learner performs an activity. If a learner answers a question correctly, the probability increases. If the learner answers incorrectly, the probability decreases.

学习者元素是学习者当前能力状态的模型。 因此，对于专家图中的每个给定节点，学习者模型都有与之相关的概率：1-99％。 每次学习者执行一项活动时，系统都会更新此图。 如果学习者正确回答了问题，则概率会增加。 如果学习者回答不正确，则概率降低。

Each learner has their own learner model. So each time there's a new learner in the system, there's a new learner model. Later, the tutor will use the learner model to decide how to order the learning content.

每个学习者都有自己的学习者模型。 因此，每当系统中有一个新的学习者时，就会有一个新的学习者模型。 稍后，导师将使用学习者模型来决定如何订购学习内容。

There's many algorithms for updating the learner model. Knowledge spaces suggest that as a learner develops a skill, the probabilities for related skills should also adjust.

有许多用于更新学习者模型的算法。 知识空间表明，随着学习者发展技能，相关技能的概率也应随之调整。

Some adaptive learning systems use simple heuristic models for updating skill probabilities. In item response theory, the probability updates along a sigmoid curve. In Bayesian knowledge tracing, this curve has a more conservative shape. Each model tends to account for these factors:

一些自适应学习系统使用简单的启发式模型来更新技能概率。 在项目响应理论中，概率沿S形曲线更新。 在贝叶斯知识跟踪中，此曲线具有更保守的形状。 每种模型都倾向于考虑以下因素：

• Before the learner does anything, what do we estimate the probability to be?在学习者做任何事情之前，我们估计概率是多少？
• How likely is a learner to guess the right answer if they don't know the skill?如果学习者不知道该技能，那么他们猜对答案的可能性有多大？
• How likely is a learner to slip up even if they know the skill?即使他们知道该技能，学习者有多大可能滑倒？
• How likely is the learner to have "learned" the skill after seeing the item?学习者在看到项目后有多大可能“学习”了技能？
• How likely does this activity categorize the learner as skilled or unskilled?此活动将学习者归为熟练或不熟练的可能性有多大？
• How difficult will this item be for this particular learner?这个项目对这个特定的学习者有多困难？

For both item response theory and Bayesian knowledge tracing, you'll need a means to estimate these parameters. This is one of the most rapidly developing areas in adaptive learning systems, so I can't make any specific recommendations yet. There's also researchers creating models with classic machine learning, such as neural networks.

对于项目响应理论和贝叶斯知识跟踪，您都需要一种估算这些参数的方法。 这是自适应学习系统中发展最快的领域之一，因此我无法提出任何具体建议。 也有研究人员使用经典的机器学习(例如神经网络)创建模型。

导师-什么时候显示 (The tutor — what to show when)

The tutor chooses in which order to select the activities which the learner will engage. After each update to the learner model, the tutor will update the path it will take to optimize for that learner.

导师选择以哪种顺序来选择学习者将从事的活动。 在对学习者模型进行每次更新之后，导师将更新为该学习者进行优化所需的路径。

The goal of the tutor is to get the learner to a complete expert graph in the smallest amount of time. Some systems allow learners to focus only on some areas while ignoring the rest. As the learner model is unique per learner, so too is the path the tutor will take. While the expert and learner elements are data with some algorithms, the tutor is algorithms with some data.

辅导者的目标是在最短的时间内使学习者掌握完整的专家图表。 有些系统允许学习者只专注于某些领域，而忽略其余部分。 由于学习者模型对于每个学习者而言都是唯一的，因此导师将采取的路径也是如此。 专家和学习者元素是具有某些算法的数据，而导师是具有某些数据的算法。

The tutor may decide both which skills to focus on and which activities to have the learner perform. For the skills to focus on, the tutor will often choose skills with the largest impact on the larger graph. This often means focusing on more elementary skills before more advanced skills. For activities:

辅导员可以决定要专注于哪些技能以及要让学习者进行哪些活动。 为了专注于技能，导师通常会选择对较大图表影响最大的技能。 这通常意味着先关注更多基础技能，然后再关注更高级技能。 活动：

• The tutor will try to choose the most relevant activities to the learner导师将尝试选择与学习者最相关的活动
• The tutor will choose activities that are challenging, but not too difficult for the learner.辅导员将选择具有挑战性但对学习者来说不太困难的活动。
• The tutor will try to choose activities in a way that reduce the total time towards mastery.导师将尝试以减少总精通时间的方式选择活动。

Simple adaptive learning tutors may choose activities within a skill at random. Item response theory based tutors emphasize choosing activities that are challenging. In Bayesian knowledge tracing models the market has many different tutor algorithms. Researchers have focused more on the expert and learner elements. So we don't know what produces the best learning outcomes for the tutor element.

简单的自适应学习导师可以随机选择一项技能内的活动。 基于项目响应理论的导师强调选择具有挑战性的活动。 在贝叶斯知识跟踪模型中，市场上有许多不同的导师算法。 研究人员将更多的精力放在专家和学习者上。 因此，我们不知道是什么因素导致了辅导元素的最佳学习效果。

界面-如何显示 (The interface — how to show it)

Some adaptive learning systems will change the user interface. As the learner is less familiar with a skill, the interface would reduce and focus more on the task at hand. As learner ability grows, more of the full interface comes together. Some call this process "scaffolding".

一些自适应学习系统将更改用户界面。 由于学习者对技能的熟悉程度较低，因此该界面将减少并更多地关注手头的任务。 随着学习者能力的提高，更多的完整界面汇集在一起​​。 有人称此过程为“脚手架”。

In some systems, learners may ask for and receive hints. When to offer hints and the depth of those hints can adjust based on learner ability.

在某些系统中，学习者可能会要求并接收提示。 何时提供提示，这些提示的深度可以根据学习者的能力进行调整。

There's also some other questions like:

还有其他一些问题，例如：

• Do you display the expert graph to the learner?您是否向学习者显示专家图？
• Do you display their progress on all skills? How?您会在所有技能上显示他们的进步吗？ 怎么样？
• Do you display their progress on specific skills? How?您是否显示他们在特定技能上的进步？ 怎么样？
• Does the learner get choices in learning content? Or does the system decide everything?学习者在学习内容上有选择吗？ 还是系统决定一切？

Depending on the needs of the system, some of these items may impact learning outcomes.

根据系统的需求，其中一些项目可能会影响学习成果。

我们如何知道自适应学习是否有好处？ (How do we know if adaptive learning is any good?)

As these systems come from academia, we have a significant amount of data and history with each system.

由于这些系统来自学术界，因此每个系统都有大量的数据和历史记录。

Human individual tutoring has the strongest learning outcomes. This is a common finding in educational research. So far, no computerized adaptive learning system has outperformed human one-on-one tutoring.

人类个人补习具有最强的学习成果。 这是教育研究中的常见发现。 到目前为止，还没有计算机化的自适应学习系统胜过人类一对一的辅导。

Researchers have investigated classroom learning alone, computerized adaptive learning alone, as well as combined classroom and adaptive learning. A 2016 paper "Effectiveness of Intelligent Tutoring Systems" provides a meta analysis of these studies. Adaptive learning systems usually outperform traditional classroom learning. Combined with classroom learning, adaptive learning systems create a positive effect, but there are some limitations.

研究人员研究了单独的课堂学习，单独的计算机自适应学习，以及课堂和自适应学习的结合。 2016年的论文“智能辅导系统的有效性”对这些研究进行了荟萃分析。 自适应学习系统通常优于传统的课堂学习。 结合课堂学习，自适应学习系统可以产生积极的效果，但存在一些局限性。

Adaptive systems do particularly well with instant feedback and ensuring skill mastery. Investigators note some areas for improvement:

自适应系统在即时反馈和确保熟练掌握方面表现特别出色。 研究人员指出了一些需要改进的地方：

• The cost of developing content for these systems is high.为这些系统开发内容的成本很高。
• These systems often can't contextualize learning the way a human can.这些系统通常无法像人类一样以上下文方式学习。
• Adaptive learning systems can feel more challenging, which can reduce learner motivation.自适应学习系统可能会更具挑战性，从而降低学习者的动机。

结语 (Wrap up)

Welp, I've nerded out now. I've covered what adaptive learning systems are. I've provided some context for the design of these systems. A touch of history. I've covered the four major elements: the expert, the learner, the tutor, and the interface. Hopefully it wasn't too technical.

抱歉，我现在已经呆呆了。 我已经介绍了什么是自适应学习系统。 我为这些系统的设计提供了一些背景信息。 一段历史。 我已经介绍了四个主要元素：专家，学习者，导师和界面。 希望它不是太技术性。

Obligatory end-of-article call-to-action: Check out Sagefy, the open-content adaptive learning system I'm working on.

强制性的号召性用语呼吁： 请查看Sagefy ，这是我正在研究的开放内容自适应学习系统。

翻译自: https://www.freecodecamp.org/news/adaptive-learning-systems/

自适应学习系统