Agents and task specifications
How should an agent act to achieve a performance measure in a task environment?
Chapters 1–2
Problem-solving · Chapters 2–4
Problems & environments, in motion
Before choosing an AI algorithm, ask two questions: what exactly is the problem, and what can the agent observe about the world? Explore classic problem examples, four problem formulations, and the properties that shape an agent’s task environment.
Chapter 4 · Search in complex environments
One task, four ways to formulate it
A classical search problem assumes a known initial state and predictable action results. Relaxing those assumptions changes what counts as a solution.
Start with the initial state, then step through a plan.
Chapter 3 · Example problems
What does the agent have to solve?
Select a classic toy or real-world problem. Each card identifies the state, goal, actions, and cost an agent would need to define before searching.
Select a problem, then animate its specification.
The example set covers standard toy problems (vacuum world, 8-puzzle, 8-queens) and common real-world formulations (route finding, traveling salesperson, robot navigation, VLSI layout, and assembly sequencing). A real application’s exact state, goal, legal actions, and cost depend on its domain.
A map of the book’s problem families
AI problems are more than search
Different chapters ask different kinds of questions: find a path, satisfy constraints, infer hidden facts, choose under uncertainty, learn from data, or control a physical system. Use this map to find a visual guide for each family.
Search, optimization, and games
Find low-cost action sequences, optimize a state, or choose against an opponent.
Chapters 3–5
Constraints and knowledge
Assign values consistently, represent facts, and derive what follows from a knowledge base.
Chapters 6–10
Planning and uncertainty
Plan action sequences, update beliefs from evidence, and decide with uncertain outcomes.
Chapters 11–18
Learning
Fit models from examples, infer latent variables, build neural networks, and learn from rewards.
Chapters 19–22
Language, vision, and robotics
Interpret and generate language, infer structure from images, and perceive and act in the physical world.
Chapters 23–26
Limits, ethics, safety, and future systems
Assess what systems can and cannot do, reason about impacts and safety, and consider how AI capabilities combine into architectures.
Chapters 27–28
Chapter 2 · The nature of environments
Classify the task environment
Environment properties are independent dimensions, not one single category. A task may be partially observable and stochastic while still being discrete and multiagent. Choose a scenario to see a typical classification and a brief reason.
Typical properties
Classifications depend on the exact formulation and agent sensors. For example, a known map can make a planning model fully observable even though navigating the physical city involves hidden traffic and changing conditions.
Observable
Fully: sensors reveal the complete relevant state. Partially: observations are incomplete or noisy; the agent may track a belief state.
Action outcomes
Deterministic: a state and action fix the next state. Stochastic: outcomes vary according to probabilities. Strategic opponents are modeled as other agents.
Structure over time
Episodic: each decision is independent. Sequential: actions affect later states and choices.
Change
Static: world stays unchanged while deliberating. Dynamic: it may change. Semidynamic: state is fixed but performance changes with time, such as a ticking clock.
State and action values
Discrete: countable states/actions. Continuous: values such as position or velocity range over intervals.
Agents and knowledge
Single-agent tasks have one decision-making agent; multiagent tasks include others whose choices matter. A known model has known rules; unknown rules must be learned or estimated.
A useful modeling checklist
PEAS describes an agent’s task: Performance measure, Environment, Actuators, Sensors. A search problem then specifies initial state, actions, transition model, goal test, and path cost. Keep the performance measure separate from path cost: they may agree, but they are not the same definition.
Explorer performance and memory
Let F be the number of formulation steps (4), P the number of problem examples (8), and T the number of environment tags in a selected scenario. Updating a formulation or example redraws O(F) or O(1) fixed-size cards; selecting a scenario builds O(T) tag and illustration elements. The page’s total interactive UI data is O(P + F + T) space. These are costs for this teaching interface—not for solving the underlying search, CSP, game, or planning problem, whose complexity depends on its state space and algorithm.