Intelligent Agents Overview

Intelligent Agents

• Agents are anything that perceives its environment through sensors and acts upon it through actuators.
• Control theory describes a closed-loop control system as a collection of mechanical or electronic devices that automatically regulate a process variable to a specific point without human interaction.
• A softbot is a software program running on a host device.
• The agent function maps all possible percept sequences to the set of formulated actions as an abstract mathematical function.
• The agent program is a concrete implementation of the function for a given physical system.
• An agent consists of architecture (hardware) and an agent program (function implementation).
• Key components of an agent include sensors, memory, and computational power.

Example: Vacuum-cleaner World

• Percepts: Location and status (e.g., [A, Dirty]).
• Actions: NoOp, Left, Right, Suck.
• Agent function: Maps percept sequences to actions.
• Example percept sequence and action: [A, Clean] → Right; [A, Dirty] → Suck
• Implemented agent program (Vacuum-Agent): Takes location and status as input and returns an action (Suck, Right, Left).
• The program prioritizes sucking if the status is Dirty and chooses to move right if the location is A and the status is clean, or moves left if the location is B and status is clean.

Rational Agents: Defining Good Behavior

• Consequentialism: Evaluates behavior based on its consequences.
• Utilitarianism: Aims to maximize happiness and well-being.
• Rational agent definition: For each possible percept sequence, the rational agent must select an action maximizing its expected performance measure according to the evidence given in the percept sequence along with internally known details.
• Performance measure: An objective criterion for agent success (often called utility function or reward function).
• Expectation: Outcome averaged over all possible situations.
• Rule: Choose the action maximizing the expected utility.

Rational Agents: Practical Considerations

• Rationality: An ideal (no one can build a perfect agent).
• Rationality ≠ Omniscience: Rational agents can make mistakes if percepts and knowledge are incomplete.
• Rationality ≠ Perfection: Rational agents maximize expected outcomes, not always actual ones.
• Rational agents explore and learn: Using percepts to complement prior knowledge and achieve autonomy.
• Rationality is bounded: By available memory, computational power, and sensors.

Environment Types

• Fully Observable: Agent's sensors give complete environmental state access.
• Partially Observable: Agent cannot see all environmental aspects (e.g., walls).
• Deterministic: Changes are entirely determined by current state and action.
• Stochastic: Changes cannot be determined from the current state and action; randomness is present.
• Known: Agent knows environmental rules to predict outcomes.
• Unknown: Outcomes cannot be predicted.
• Static: Environment doesn't change while the agent deliberates;
• Dynamic: Environment changes during deliberation.
• Discrete: Environment has a fixed number of percepts, actions, and states;
• Continuous: Percepts, actions, and states are infinite in number;
• Episodic: Agent's actions in one episode don't affect subsequent episodes;
• Sequential: Agent's actions affect future outcomes;
• Single agent: Agent operates by itself.
• Multi-agent: Agents cooperate or compete in the same environment.

Agent Hierarchy

• Simple reflex agents: Agents react to percepts without considering past information.
• Model-based reflex agents: Maintain internal state for better decisions.
• Goal-based agents: Actions are aimed at achieving a particular goal.
• Utility-based agents: Actions are chosen to maximize expected utility.

Designing a Rational Agent

• Agent has to understand its task definition.
• Agent designs the process by which it will sense data input.
• Agent must understand what actions it takes to achieve its objective.
• A rational agent continuously assesses its performance and adjusts its actions accordingly.

Modern Vacuum Robot Example

• Features: Control via app, cleaning modes, mapping, navigation, and boundary blockers.
• Performance measure: Time to clean (95%), avoiding getting stuck.
• Environment: Rooms, obstacles, dirt, people, pets.
• Actuators: Wheels, brushes, blower, and sound (communicate instructions to server).
• Sensors: Bumpers, cameras, dirt sensors, laser, motor sensors, cliff detection, home base locator.

Intelligent Systems: Self-driving Car

• High-level planning: Designing passenger journey with an enjoyable drive.
• Low-level planning: Reactions to real-time incidents like children running in front of the car. Agents respond efficiently when unexpected events emerge.
• Agent function maps sensor data and internal state into an immediate action.

AI Areas

• Search: Finding goals like navigation.
• Optimization: Maximizing objectives like utility.
• Constraint satisfaction: Keeping within limitations like battery power.
• Uncertainty: Acknowledging and dealing with uncertain situations such as traffic flow.
• Sensing: Including language processing and vision.

What You Should Know

• Agent function: Describes how an agent interacts with its environment.
• Transition Function: Explains how the environment changes based on agent actions.
• States: Different states within the environment.
• Environment differences: Observability, uncertainty, and known vs unknown transition functions.
• Agent Types: Distinguishing diverse agent types and their specifications.

This quiz covers the fundamentals of intelligent agents, focusing on their architecture, functions, and examples like the vacuum-cleaner world. Explore how these agents perceive their environment and make decisions based on their sensors and actuators. Ideal for students studying artificial intelligence and robotics.