is the capability of computational systems to perform tasks associated with human intelligence, using learning and reasoning to achieve defined goals.

Context: Tests AI definition and scope; also checks whether the student can distinguish AGI from the general AI definition.

In AI history, an is a period of disappointment and loss of funding in AI history.

Context: Tests AI history concept: AI winter.

Step-by-step reasoning algorithms scale poorly with problem size, which causes and leads to failure on large reasoning problems.

Context: Tests cause→effect relationship using combinatorial explosion.

Uncertainty about the current situation or action outcomes causes the agent to make probabilistic guesses and reassess after acting, because non-deterministic environments require decision-making and updating based on results.

Context: Tests cause→effect mechanism: uncertainty leads to probabilistic decision-making.

A encodes concepts and relationships so programs can make deductions and answer questions using a knowledge base and formal structures like ontologies.

Context: Tests the core concept: knowledge representation and ontologies.

A is a body of knowledge represented in a form that a program can use.

Context: Tests distinction: knowledge base as the stored body of knowledge.

An selects actions to make them happen by maximizing preferences or goals.

Context: Tests core concept: rational agent.

In automated decision-making, expected utility is computed by weighting outcome utilities by their probabilities for a chosen action.

Context: Tests expected utility definition details.

A is a decision model with a transition model and reward function, plus a policy mapping states to decisions.

Context: Tests core concept: Markov decision process (MDP).

Local search via minimizes a loss function, which allows neural networks to be trained by iteratively improving parameters.

Context: Tests search/optimization technique: gradient descent.

In gradient descent, parameters are adjusted to minimize a loss function; backpropagation is used to train neural networks by computing so parameter updates reduce prediction error.

Context: Tests mechanism: backpropagation computes gradients.

is learning where training data includes labeled expected answers.

Context: Tests ML paradigm: supervised learning.

is learning where the model finds patterns or structures in unlabeled data.

Context: Tests ML paradigm: unsupervised learning.

The architecture is a deep learning architecture using an attention mechanism, enabling modern NLP and generative models.

Context: Tests core concept: transformer architecture.

Using GPUs to accelerate neural networks causes deep learning to outperform previous AI techniques, which leads to increased interest after 2012.

Context: Tests cause→effect relationship: GPUs → deep learning outperforms → interest increases.