Lecture 2: Intelligent Agents

Lecture 2: Intelligent Agents¶

AIMA Chapter 2 — 1 hour¶

Learning Objectives¶

• Define agents and environments

• Understand rationality and performance measures

• Classify task environments by key properties

• Describe agent architectures: reflex, model-based, goal-based, utility-based, learning

Agents and Environments¶

• Agent: Entity that perceives and acts

• Environment: Everything outside the agent

• Percept: Agent’s perceptual input at any instant

• Percept sequence: Complete history of percepts

The Agent-Environment Loop¶

• Agent receives percepts, chooses actions

• Environment changes in response to actions

• Agent’s next percept depends on the new state

Good Behavior: Rationality¶

• Rational agent: For each possible percept sequence, selects action expected to maximize performance measure

• Performance measure: Criterion for success (e.g., score, safety)

• Omniscience: Knowing actual outcome of actions — we assume agents don’t have this

Rationality vs. Omniscience¶

• Rational ≠ omniscient

• Rational agent acts on available information

• Learning: Improve from experience

• Autonomy: Operate correctly without constant human intervention

Specifying the Task Environment¶

PEAS framework:

• Performance measure

• Environment

• Actuators

• Sensors

Example: Autonomous taxi

• P: Safe, fast, legal, comfortable

• E: Streets, traffic, pedestrians

• A: Steering, accelerator, brake, display

• S: Cameras, sonar, GPS, odometer

Properties of Task Environments¶

Fully vs. Partially Observable¶

• Fully observable: Sensors give access to complete state

• Partially observable: Noisy/incomplete sensors (e.g., poker, medical diagnosis)

• Unobservable: No sensors — agent must act blindly

Single vs. Multi-Agent¶

• Single-agent: Only one agent (e.g., crossword puzzle)

• Multi-agent: Other agents (competitive or cooperative)

• Competitive: Zero-sum games

• Cooperative: Team goals

The Structure of Agents¶

• Agent program: Implementation of agent function

• Agent function: Maps percept sequences to actions

• Agent = architecture + program

Simple Reflex Agents¶

• Select action based on current percept only

• Condition–action rules: if percept then action

• No memory of past percepts

• Limitation: Cannot handle partial observability

Model-Based Reflex Agents¶

• Maintain internal state to track world

• Model: How world evolves, effect of actions

• Update state from: (1) how world changes, (2) how actions affect world

• Can handle partially observable environments

Goal-Based Agents¶

• Have goals — desirable states

• Consider future: “What if I do action A?”

• Search and planning to achieve goals

• More flexible than reflex agents

Utility-Based Agents¶

• Utility function: Maps states to real numbers (degree of happiness)

• Handles trade-offs (e.g., fast vs. safe)

• Handles uncertainty (expected utility)

• Generalizes goal-based: goal = states with utility above threshold

Learning Agents¶

• Performance element: Selects actions (like previous agents)

• Learning element: Modifies performance element

• Critic: Feedback on how well agent is doing

• Problem generator: Suggests exploratory actions

Summary: Agent Types¶

References¶

• Russell & Norvig, AIMA 4e, Ch. 2

• Chapter PDF: chapters/chapter-02.pdf

• Exercises: aimacode.github.io/aima-exercises/

Questions?¶

Next lecture: Solving Problems by Searching (Chapter 3)