Concept Glossary: R¶

This page lists English glossary entries for this letter. Entry bodies are generated from per-term source files with includes.

random forest¶

Meaning: An ensemble model that trains many decision trees with controlled randomness and combines their predictions to produce a more stable result.
Why it matters: A single decision tree is easy to read but can be sensitive to accidental patterns in the training data. Random forest is a representative way to reduce that instability by aggregating many trees.
Related concepts: decision tree, ensemble, bootstrap, OOB score, feature importance
Core Section: P4-15.1
Appears in: P4-15.2, P4-15.3, P4-15.4

ranking¶

Meaning: Ranking is a problem type that orders multiple candidates by score or priority. Instead of immediately choosing one correct answer, it rearranges candidates by what should be shown or checked first.
Why it matters: Search results, recommendation lists, document candidates, and review queues often care more about whether the top few candidates are useful than about each candidate's absolute score. Ranking separates how large is the score? from which candidate should come first?
Related concepts: recommendation task, search, generation
Core Section: P1-13.2
Appears in: P1-1.2, P3-9.8, P3-9.13, P4-6.4

rate of change¶

Meaning: Rate of change compares how much an output changes when an input changes. It asks how quickly or sensitively a value moves, not only how large the value is.
Why it matters: Average rate of change, derivatives, and gradients all come from the same change question. This concept prepares readers to interpret loss reduction, function sensitivity, and gradient direction.
Related concepts: derivative, gradient, limit
Core Section: P2-4.2
Appears in: P2-4.3, P2-4.4

recall¶

Meaning: Recall is the share of actual positive cases that the model successfully catches as positive.
Why it matters: Recall is important when false negatives are costly. It helps reveal whether a model is missing cases that should have been found, especially when important positives are rare.
Related concepts: precision, accuracy, confusion matrix, false negative
Core Section: P1-13.4
Appears in: P4-4.1, P4-4.2, P4-6.1, P4-6.4, P4-8.1, P4-8.2, P4-8.3

recommendation task¶

Meaning: A recommendation task is a modeling task that chooses which candidates to show or suggest using user, item, and context information. It is often about ordering useful candidates rather than finding one single correct answer.
Why it matters: A recommendation task combines prediction and ranking. It helps distinguish search, recommendation, advertising, and feed ordering even when they all present candidates to users.
Related concepts: ranking, generation, prediction
Core Section: P1-1.2

regression¶

Meaning: Regression is a modeling task that predicts a continuous numeric value or score from inputs. Unlike classification, which chooses a category name, regression estimates a value on a numeric axis such as price, temperature, or demand.
Why it matters: Many real problems ask how much? rather than which category? Regression helps separate classification-style right/wrong evaluation from distance-based error evaluation, and it explains why accuracy is not the right metric for continuous outputs.
Related concepts: classification, model output, task definition
Core Section: P1-8.1
Appears in: P1-4.4, P1-6.3, P3-9.13, P4-2.1, P4-3.2, P4-6.2, P4-8.1, P4-8.3, P4-10.1

regularization¶

Meaning: Regularization is the idea of adding constraints or extra costs during learning so a model does not fit only the training data too aggressively. Instead of rewarding only lower training error, the training criterion also makes overly complex or unstable solutions less attractive. Temporarily dropping some node outputs or connections during training is one representative regularization strategy.
Why it matters: A model can score well on training data and still fail on new data. Regularization makes learning about both fitting the data and avoiding unnecessarily complex solutions. For example, discouraging very large weights or overly sensitive patterns may slightly reduce training score while improving generalization. This keeps best training score from being treated as the same thing as best model.
Related concepts: overfitting, generalization, model validation
Core Section: P5-8.1
Appears in: P4-5.1, P4-11.5, P5-8.2

reinforcement learning¶

Meaning: Reinforcement learning is a learning setup where an agent chooses actions in states and learns a better policy from rewards returned after those actions.
Why it matters: Reinforcement learning separates sequential decision problems from supervised labels and unsupervised structure discovery. It asks not only whether one immediate answer is correct, but how a series of actions affects long-term reward.
Related concepts: reward, reinforcement learning policy, reinforcement learning agent
Core Section: P4-2.3
Appears in: P1-8.3, P4-index, P4-2.3, P4-3.2, P4-19.1, P4-19.2, P4-19.3, P4-19.4

reinforcement learning agent¶

Meaning: In reinforcement learning, a reinforcement learning agent is the learning subject that chooses actions from the current state or observation, receives rewards, and updates its policy. It is not just an executor. It interacts with an environment, accumulates experience, and changes the basis for later action.
Why it matters: This term separates learning through action outcomes inside an environment from service AI agents that connect tool calls and task workflows. The same general word agent can appear in both contexts, but the learning role, operating layer, and failure modes are different.
Related concepts: reinforcement learning, action, reward
Core Section: P1-8.3
Appears in: P1-7.1

reinforcement learning policy¶

Meaning: A reinforcement learning policy is the rule or function that chooses an action from the current state or observation. It may be a direct mapping from situations to actions, or a rule that assigns probabilities to possible actions.
Why it matters: Reinforcement learning is not about guessing one answer once. It is about learning a way to choose actions repeatedly across changing states. A policy helps distinguish whether a system got one step right from whether it keeps making useful choices over many steps.
Related concepts: reinforcement learning, state, action
Core Section: P1-8.3
Appears in: P1-14.3, P4-2.3, P4-11.1, P4-19.1, P4-19.2

rule-based approach¶

Meaning: A rule-based approach solves a problem by having people explicitly write conditions, facts, procedures, or policies and then applying those rules to reach a judgment or action.
Why it matters: It gives a clear contrast with machine learning. In a rule-based approach, the judgment criterion is written directly; in a learning-based approach, the criterion is fitted from data cases.
Related concepts: rule-based system, knowledge representation, machine learning, policy rule
First explained in: P1-2.1
Appears in: P4-1.2

representation¶

Meaning: An internal form that makes input usable by a model. It is not the raw data itself, but a computational form reorganized so useful differences and relationships for the current task become easier to use.
Why it matters: To understand deep learning and representation learning, the reader must separate the input that humans see from the internal form a model computes with. If a feature is an input clue, a representation is closer to how those clues are reorganized inside the model. A good representation can make later classification or generation easier.
Related concepts: feature, vector, activation, embedding, representation learning
Core Section: P1-3.3
Appears in: P1-2.3, P1-4.3, P1-8.2, P3-6.1, P3-6.2, P4-13.2, P5-2.2, P5-10.1, P5-10.2

representation learning¶

Meaning: An approach in which a model learns how to turn input into a more useful internal representation from data. Instead of having people predefine every rule and feature, the model learns which representation helps the current task. In that sense, representation learning is closer to the model learns a useful internal coordinate system from data than to people choose all good features in advance.
Why it matters: This concept separates human-designed features from the deep learning step that learns internal representations. It helps readers see that deep learning is not only about adding more layers, but also about learning intermediate representations that can make later classification, prediction, or generation easier. It also makes the boundary in Part 3 clearer: feature design specifies the input structure first, while representation learning happens later inside the model.
Related concepts: deep learning, feature, representation
Core Section: P1-3.3
Appears in: P1-2.3, P1-9.1, P3-6.3, P5-2.2, P5-10.1, P5-10.2, P6-3.3

topology¶

Meaning: Topology is a mathematical viewpoint that studies the structure of a space through ideas such as open sets, neighborhoods, connectedness, and continuity, rather than through coordinates or distances themselves. If position in P2-3.2 means the coordinate-like place of one vector, topology looks more broadly at the connected and continuous structure of the whole space. It is not the same as a single position.
Why it matters: In AI writing, phrases such as topology of a representation space or data manifold often refer to the overall structure made by data representations rather than individual vector coordinates. This distinction prevents beginners from collapsing position, distance, and topology into one vague space metaphor. Here, the standard mathematical term topology is only introduced briefly in the context of representation spaces; this glossary entry does not develop a full topology course.
Related concepts: vector space, dimension, representation
Core Section: P2-3.2

reproducibility¶

Meaning: Reproducibility is the ability to obtain similar results again when the same code, data, settings, and procedure are prepared. It is not a one-time successful run; it is a documented condition that can be inspected and repeated.
Why it matters: Learning examples, experiments, and project reviews need more than a code file. Reproducibility keeps data versions, execution order, configuration, evaluation procedure, and records connected so later comparison and debugging remain meaningful.
Related concepts: standardization, model validation, data leakage
Core Section: P2-7.5
Appears in: P1-12.3, P1-14.5, P1-14.6, P3-9.7, P7-3.2, P7-7.2

retrieval-augmented generation, RAG¶

Meaning: Retrieval-augmented generation, or RAG, is a structure that first retrieves external material related to a question, attaches that material to the model input context, and then generates an answer. The key point is that retrieval and generation are separate stages, letting the model use both what it learned in its parameters and the evidence just retrieved.
Why it matters: RAG is a representative way to reduce the limits of answering only from internal model parameters and to ground responses in current material or organization documents. It also shows that answer quality can fail if retrieval quality, chunk design, or source tracking is weak, even when the generation step sounds fluent.
Related concepts: embedding, similarity search, retrieval, provenance, search index
Core Section: P1-13.3
Appears in: P1-14.1, P1-14.2, P1-14.3, P6-4.2, P6-10.3, P6-12.1, P6-12.2, P6-13.1, P7-5.1, P7-5.3

RAG retrieval¶

Meaning: RAG retrieval is the stage that brings in external documents or chunks likely to be relevant to the user's question. It should be read more narrowly than general search or a person's web browsing: it prepares candidate evidence to be placed into the generation input.
Why it matters: The word retrieval can point to state-space search, document search, web search, or retrieval in RAG. In RAG, retrieval does not directly create the answer; it selects outside material that the generative model may use as additional input context. This distinction makes clear that retrieved results are not automatically the answer.
Related concepts: retrieval-augmented generation, RAG, information retrieval, generation
Core Section: P1-13.3
Appears in:

reward¶

Meaning: A reward is the feedback signal returned after an action to show how much that result helped the goal. In reinforcement learning, the environment summarizes the usefulness of an action as a number.
Why it matters: A reward is not the same as a supervised-learning answer label. It appears after action and guides policy adjustment through immediate and long-term results. Understanding reward also shows why poor reward design can teach behavior that optimizes the number while missing the real objective.
Related concepts: reinforcement learning, reinforcement learning policy, reinforcement learning environment
Core Section: P1-8.3
Appears in: P4-index, P4-2.3, P4-19.1, P4-19.2, P4-19.3, P4-19.5, P4-summary

reward design¶

Meaning: Reward design is the work of defining, through a reward function and constraints, what a reinforcement-learning agent should treat as doing well. It is a form of goal definition, not only a numeric implementation detail.
Why it matters: If the reward only roughly stands in for the real goal, the AI agent can raise the number while missing the human intention. Reward design is therefore the starting point for checking reward hacking, safety, and deployment risk.
Related concepts: reward, true objective, reward hacking, proxy target
Core Section: P4-19.3
Appears in:

reward hacking¶

Meaning: Reward hacking is the situation where a model or reinforcement learning reinforcement learning agent optimizes the reward function literally while missing the meaning humans intended, raising the number without achieving the real goal.
Why it matters: Reinforcement learning directly maximizes reward, so a mismatch between reward and objective can be optimized aggressively. This concept separates the score increased from the goal was achieved.
Related concepts: reward design, true objective, proxy target, AI alignment
Core Section: P4-19.3
Appears in: P4-19.4

RNN, recurrent neural network¶

Meaning: An RNN is a neural-network structure that passes a hidden state from one step to the next while processing ordered data. It reads a sequence one step at a time and accumulates information in an internal state.
Why it matters: RNNs show an early neural-network answer to the question of how order can be remembered in language, time series, and other sequential data. Understanding RNNs also makes it easier to compare why Transformers later handled long-context relationships differently.
Related concepts: hidden state, long-term dependency, sequence modeling
Core Section: P1-11.2
Appears in: P5-12.1, P5-12.2

rule-based system¶

Meaning: A rule-based system compares current facts with human-written rules to decide a conclusion or action. It applies explicit if this, then that criteria rather than learning numerical boundaries from data.
Why it matters: Rule-based systems clarify the strengths and limits of explainable, controllable, policy-like automation. They work well when rules are explicit, but they can become brittle when exceptions are numerous or boundaries are unclear.
Related concepts: symbolic AI, inference engine
Core Section: P1-3.1
Appears in: P1-2.1

rightsholder¶

Meaning: A person or organization that owns or controls rights in a work or protected material.
Why it matters: It matters when deciding permission, licensing, attribution, and reuse boundaries.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-15.2

RLHF, reinforcement learning from human feedback¶

Meaning: A training approach that uses human preference feedback as a signal for adjusting model behavior.
Why it matters: It is important in LLM alignment, but it should not be confused with all reinforcement learning.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-8.3

reasoning¶

Meaning: The process of connecting premises, rules, evidence, or intermediate steps to reach a conclusion.
Why it matters: It must be distinguished from model inference, prediction, and generation, which can be different technical processes.
Related concepts: Read together with the surrounding glossary entries and the linked Sections for this term.
Core Section: P1-5.3