John McCarthy coined the term “Artificial Intelligence” in 1956 and organised the Dartmouth Conference, which is why he is regarded as the Father of AI.

The core aim of AI is to build machines that can perceive, reason, learn and make decisions in a way that simulates human intelligence.

The Turing Test, proposed by Alan Turing, judges a machine’s ability to produce responses indistinguishable from those of a human.

IPL (Information Processing Language), developed in 1956 by Newell, Shaw and Simon, was the earliest language created specifically for AI research — predating LISP (1958) and PROLOG.

Narrow (Weak) AI is built for a single, well-defined task such as facial recognition or playing chess, and cannot generalise beyond it.

An agent perceives its environment through sensors and acts upon that environment through actuators.

DFS only needs to store the nodes along the current path (linear space), whereas BFS keeps every generated node in memory.

Hill Climbing greedily moves to a better neighbour, so it can easily get trapped at a local maximum (or on a plateau) instead of reaching the global optimum.

Artificial Neural Networks are loosely modelled on biological neurons, using weighted connections and activation functions to learn patterns.

Reinforcement Learning trains an agent through rewards and penalties received from interacting with its environment.

A* is a refined form of Best-First Search: it evaluates nodes with f(n) = g(n) + h(n), combining path cost with a heuristic estimate.

Alpha-Beta Pruning discards branches that provably cannot influence the final minimax decision, letting the same search run deeper in the same time.

A simple reflex agent ignores history and reacts purely to the current percept using condition–action rules.

NLP stands for Natural Language Processing — the AI subfield concerned with enabling machines to understand and generate human language.

K-Means partitions unlabeled data into k clusters by iteratively minimising the distance of points to their cluster centroid.

Computer Vision enables machines to acquire, process and interpret images and video to support decision-making.

Probability theory and Bayesian Networks model uncertainty by representing degrees of belief and conditional dependencies among variables.

An Expert System encodes a human specialist’s knowledge as rules in a knowledge base and reasons over them with an inference engine.

Problem formulation begins with the Initial State — the state in which the agent starts before taking any action.

Overfitting means the model has memorised the training data (including its noise) and therefore generalises poorly to unseen data.

In his 1950 paper, Turing called it the “Imitation Game” — a machine tries to imitate human responses convincingly.

Greedy Best-First Search always expands the node with the smallest heuristic value h(n), i.e. the one that appears closest to the goal.

Activation functions add non-linearity, which allows the network to learn complex, non-linear mappings between inputs and outputs.

Natural Language Processing deals with analysing and generating human language, including meaning, syntax and semantics.

An agent is the combination of its architecture (the computing platform with sensors and actuators) and its program (the agent function).

The universal quantifier ∀ means “for all”, while ∃ is the existential quantifier meaning “there exists”.

Supervised Learning trains on labelled input–output pairs, comparing predictions with the known correct outputs.

In a Fully Observable environment, the agent’s sensors give it access to the complete state of the environment at each moment.

The Minimax algorithm assumes an optimal opponent and chooses the move that minimises the maximum possible loss.

The Knowledge Base stores facts and rules about the world; the inference engine derives new conclusions from them.

Hyperparameter Tuning searches for the best settings (learning rate, depth, number of units, etc.) that maximise validation performance.

Breadth-First Search is uninformed — it uses no heuristic information about how close a node is to the goal.

A Model-Based Reflex Agent keeps an internal model/state so it can handle partially observable environments.

Transfer Learning reuses the knowledge stored in a pre-trained model and fine-tunes it for a new but related task.

Fuzzy Logic replaces crisp true/false with a continuous degree of membership between 0 and 1.

Internal nodes test attributes; leaf (terminal) nodes hold the final predicted class label or decision.

A loss function quantifies the error — how far the model’s predictions are from the true target values — and drives parameter updates.

BFS expands level by level, so the first goal node it reaches is guaranteed to be the shallowest one.

Regularization (L1/L2) adds a penalty on large weights to the loss, discouraging the model from fitting noise.

AGI stands for Artificial General Intelligence — a hypothetical AI with human-level, general-purpose reasoning across many tasks.

The Sigmoid function squashes its input into the open interval (0, 1). Tanh, by contrast, ranges from −1 to 1.

The Inference Engine applies logical rules to the facts in the Knowledge Base to derive new conclusions.

SVMs use kernel functions to map data into a higher-dimensional space where a maximum-margin hyperplane separates the classes.

A Confusion Matrix tabulates TP, FP, TN and FN, and from it metrics such as accuracy, precision and recall are derived.

Gradient Descent iteratively moves parameters in the negative gradient direction to reduce the loss.

Backpropagation applies the chain rule backwards through the network to compute the gradient of the error w.r.t. each weight.

The Transformer architecture, introduced in the 2017 paper “Attention Is All You Need”, is built entirely on self-attention.

The “naive” assumption is that all features are conditionally independent of one another given the class label.

The validation set is held out from training and used to tune hyperparameters and detect overfitting; the test set is used for final evaluation.

Backtracking Search assigns variables one at a time and undoes assignments that violate constraints — the standard CSP solver.

The discount factor γ (0 to 1) determines how heavily future rewards are weighted relative to immediate ones.

RNNs maintain a hidden state that carries information across time steps, making them suitable for sequences.

In a GAN, the Generator tries to produce realistic data while the Discriminator tries to tell real from fake, and both improve through competition.

PCA projects data onto new orthogonal axes (principal components) that capture the greatest variance.

The percept sequence is the full history of percepts the agent has received; the agent function maps it to an action.

Pooling layers (e.g. max or average pooling) reduce the spatial resolution of feature maps, lowering computation and adding translation invariance.

Dropout randomly switches off a proportion of neurons during each training step, preventing the network from over-relying on any single unit.

Random Forest is an ensemble of decision trees trained on bootstrapped samples and random feature subsets, then averaged or voted.

High bias causes underfitting; high variance causes overfitting. The goal is to find the complexity that minimises total error.

An action’s Effect describes how the world changes when the action is executed; its Precondition must hold before it can be applied.

Uniform Cost Search always expands the node with the smallest accumulated path cost g(n), making it optimal for non-negative step costs.

A* evaluates nodes with f(n) = g(n) + h(n), combining the cost so far with the estimated cost to the goal.

An admissible heuristic never overestimates the true cost to the goal, which guarantees A* finds an optimal solution.

Exploration tries new or random actions to gather information, whereas exploitation uses the best-known action.

The agent function specifies which action the agent takes for every possible percept sequence.

A dynamic environment can change while the agent is still deciding what to do; a static one only changes when the agent acts.

Text-to-Speech (TTS) synthesis converts written text into audible speech; speech recognition does the reverse.

Optical Character Recognition extracts machine-readable text from images of printed, typed or handwritten text.

Accuracy = (TP + TN) / total predictions — the proportion of all predictions that were correct.

The F1-Score is the harmonic mean of Precision and Recall, giving a balanced single-number measure.

Hill Climbing keeps only one current state and repeatedly moves to a neighbouring state with a better value.

The Softmax layer exponentiates and normalises the logits so they sum to 1, producing a probability for each class.

SMOTE creates new synthetic minority-class examples by interpolating between existing minority instances and their nearest neighbours.

Simulated Annealing borrows the annealing metaphor: a “temperature” parameter allows occasional worse moves early on, gradually cooling to converge.

Genetic Algorithms evolve a population of candidate solutions using selection, crossover and mutation operators.

The agent must balance exploring new actions to learn more against exploiting known good actions to maximise reward.

Regression predicts continuous numeric outputs such as prices, temperatures or demand.

Classification assigns each input to one of a finite set of discrete class labels.

Explainable AI (XAI) aims to make model decisions transparent and interpretable to humans.

In a stochastic (non-deterministic) environment, the same action can lead to different outcomes.

Historical data can encode past discrimination, so models trained on it may reproduce or amplify algorithmic bias.

Naive Bayes applies Bayes’ Theorem while naively assuming that all features are conditionally independent given the class.