Juniper Mist Cloud Fundamentals · Topic 1 of 2

AI Concepts and Machine Learning

Learn AI and machine learning for the JN0-253: supervised and reinforcement learning, mutual information, decision trees, LSTM and NLP in Mist AI.

  • 6 min
In a nutshell
  • AI is the broad goal, ML learns from data, deep learning uses layered neural networks.
  • Mist AI is data plus AI primitives plus a virtual assistant (Marvis).
  • Supervised learning uses labeled data; unsupervised finds patterns in unlabeled data; reinforcement learns from rewards by trial and error.

What it is

Artificial intelligence (AI) is the broad goal of making software do tasks that normally need human judgment, such as spotting a problem, explaining its cause or answering a question. Machine learning (ML) is the part of AI where the software learns from data instead of following hand-written rules. Deep learning is a part of ML that uses many-layered neural networks.

Juniper describes Mist AI as a toolbox for a self-driving network. It combines three things: data collected from the network, AI primitives built with data science tools, and a virtual assistant, Marvis, that people can talk to.

Artificial intelligencecontainsMachine learningcontainsDeep learningneural networks

How it works

The three learning types

Machine learning models are trained in one of three main ways.

Learning typeWhat it learns fromTypical resultMist example
SupervisedLabeled examples with a known correct outcomeClassifies new cases or predicts a valueDecision trees that classify issues such as a faulty cable
UnsupervisedUnlabeled data, with no answer givenGroups, patterns and outliersNot one of the named Mist tools
ReinforcementRewards and penalties from its own actionsA policy: the best action in each situationRadio resource management (RRM)

Supervised learning needs labels. A model is shown many examples where the outcome is already known, for example cable tests marked as good or bad, and learns which input values lead to which outcome.

Unsupervised learning has no labels; it finds clusters, patterns and unusual values on its own.

Reinforcement learning uses an agent that takes an action, sees the result and receives a reward or a penalty. It learns by trial and error, favoring the actions that earn rewards.

The Mist data science toolbox

Juniper names a set of tools that range from simple statistics to deep learning.

ToolKindWhat Mist uses it for
Mutual informationStatistics, based on entropyFinding which network feature tells you most about an SLE pass or fail
Pearson correlationStatisticsShowing whether a feature points toward failure (negative) or success (positive)
Decision treesSupervised learningPinpointing issues such as faulty cables, MTU mismatch, incomplete negotiation, AP or switch health and coverage
LSTM recurrent neural networkDeep learningUsing earlier network events to judge what is happening now
Reinforcement learningAgent learning from rewardsTuning radio settings per site with RRM
Natural language processing (NLP)Language modelsLetting Marvis understand questions typed in plain English

Mutual information and SLEs

A service level expectation (SLE) metric such as Time to Connect gives a pass or fail for each user minute. Mutual information measures how much a feature, such as device type, operating system, user, AP or time interval, reduces the uncertainty about that pass or fail.

Entropy is the measure of uncertainty. If a feature tells you nothing about the SLE metric, the mutual information is zero. A feature with high mutual information is a strong suspect when the SLE fails.

SLE failuresmutual informationRanked featuresdevice type, AP, OScorrelation signLikely cause

Decision trees

A decision tree splits data with a series of yes or no questions until the labels are sorted with high certainty. Each split is chosen with Gini impurity, the uncertainty at a node, and information gain, how much a split reduces that uncertainty.

Pruning removes branches to stop the tree from fitting noise in the training data, called overfitting. Juniper lists simple decision trees, random forest, gradient boosting and XGBoost among the tree methods.

Reinforcement learning for RRM

Mist uses reinforcement learning to tune the radio environment of each site. The agent's value function is based on coverage, capacity and connectivity.

RRM action the agent can takeEffect
Switch between 2.4 GHz and 5 GHzChanges the band a radio serves
Adjust transmit powerGrows or shrinks the cell
Change channelMoves away from interference
Adjust channel bandwidthTrades throughput against channel reuse
Switch BSS color (Wi-Fi 6 APs)Reduces co-channel contention

The learned sequence of actions forms a policy. Because every site learns its own policy, each site ends up with RF settings fitted to that building.

RRM agenttakes actionRadio changepower, channel, bandmeasuresCoveragecapacity, connectivityrewardRRM agent

NLP and Marvis

Natural language processing gives software the ability to derive meaning from human language. For Marvis, a question is cleaned of stop words, split into tokens, turned into vectors and compared with known questions.

Training data comes from real customer questions annotated with intents, such as troubleshoot or count, and entities, such as a device name or a time frame. For a new question, the model finds the most similar known intent and entities. Juniper also states that Marvis now draws on large language models and generative AI.

Exam tip

When a question asks which technique lets Marvis answer "Why did this user have a bad connection yesterday?", the answer is natural language processing. When it asks how Mist decides which feature explains SLE failures, the answer is mutual information.

Common mistakes

Mixing up supervised and reinforcement learning. Supervised learning needs labeled outcomes in advance. Reinforcement learning has no labeled answers; it learns from the reward it receives after acting.

Calling all of Mist AI "deep learning". Mist uses a range of tools. Mutual information and correlation are statistics, decision trees are classic supervised ML, and only the LSTM networks are deep learning.

Reading a high mutual information value as "cause proven". Mutual information shows that a feature is informative about failures. It points to the likely cause; it does not prove it alone.

Watch out

Do not confuse the RRM reward inputs with SLE names. The RRM value function uses coverage, capacity and connectivity; the SLE list is a separate monitoring feature covered in a later chapter.

What to remember for the exam

  • AI is the broad goal, ML learns from data, deep learning uses layered neural networks.
  • Mist AI is data plus AI primitives plus a virtual assistant (Marvis).
  • Supervised learning uses labeled data; unsupervised finds patterns in unlabeled data; reinforcement learns from rewards by trial and error.
  • Mutual information, based on entropy, ranks which features explain SLE failures; zero means the feature tells you nothing.
  • Decision trees use Gini impurity and information gain; pruning limits overfitting.
  • LSTM is a recurrent neural network that uses earlier events to judge current ones.
  • RRM uses reinforcement learning: band, power, channel, bandwidth and BSS color actions, rewarded on coverage, capacity and connectivity.
  • Marvis uses NLP with intents and entities to understand plain-language questions.

Check yourself

  • Which learning type does Mist use for radio resource management?

    Reinforcement learning, where an agent changes radio settings and is rewarded on coverage, capacity and connectivity.

  • What does a mutual information value of zero between a feature and an SLE metric mean?

    The feature tells you nothing about whether the SLE passes or fails, so it is not a useful suspect.

  • What are intents and entities in Marvis NLP?

    An intent is what the user wants to do, such as troubleshoot or count, and an entity is a detail in the question, such as a device name or time frame.

Topic 1 of 2 in Juniper Mist Cloud Fundamentals

Test yourself on Juniper Mist Cloud Fundamentals

The full Juniper JNCIA-MistAI JN0-253 question bank has 668 questions, with timed practice exams and an explanation for every answer.

This study guide is independent and is not affiliated with, endorsed by, or sponsored by Juniper. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.