Intelligent Agents

This outlines the fundamental concepts of how agents interact with their environments, the criteria for rationality, and the various architectures used to build intelligent systems.

1. Foundations of Agents and Environments

An agent is anything that can perceive its environment through sensors and act upon that environment through actuators

• Percept: Any input the agent receives at a given moment (e.g., a driving agent perceiving a pedestrian or a traffic light)

• Percept Sequence: The complete history of everything the agent has perceived

• Agent Function/Program: The internal logic that maps a specific percept sequence to an allowed action

2. Rationality and PEAS

A rational agent is one that performs the "right" sequence of actions to maximize its performance measure (rewards)

• Rewards: These are designed based on domain expertise to ensure the agent achieves long-term goals

. For instance, a vacuum cleaning agent should be rewarded for keeping a room clean while also minimizing electricity consumption

• PEAS Framework: To design a rational agent, one must specify its Performance measure, Environment, Actuators, and Sensors

. For example, a "Medical diagnosis system" has a performance measure of "Healthy patient, reduced costs," while its sensors include "Keyboard entry of symptoms"

3. Properties of Task Environments

Environments are categorized by several key parameters that dictate how complex an agent must be

• Observability: Can be Fully Observable (face recognition), Partially Observable (driving), or Unobservable (path navigation)

• Number of Agents: Can be Single agent (crossword solver) or Multi-agent, which is further split into Competitive (chess) or Cooperative (rescue robots)

• Certainty: Deterministic environments have predictable outcomes (Sudoku), while Stochastic environments involve uncertainty (medical robots)

• State: Can be Discrete (N-queens) or Continuous (fire alarm)

• Nature: Known environments have fixed rules, whereas Unknown environments require the agent to learn the rules (e.g., a taxi driver in a new country)

• Episodes: Episodic tasks are independent of each other (biometric recognizer), while Sequential tasks require the agent to consider the impact of current actions on future states (ICU monitor).

4. Types of Agent Architectures

There are five main types of agents, ranging from simple to highly complex

• Simple Reflex Agents: These provide instantaneous responses based on current percepts using "condition-action rules" (e.g., closing your eyes when dust approaches)

• Model-based Reflex Agents: These maintain a "mental model" of how the world evolves to prepare for future actions, even if those actions aren't immediately triggered by current sensors (e.g., preparing for a toll booth after seeing a sign 100m away)

• Goal-based Agents: These agents act to achieve a specific destination or state, requiring planning and search capabilities (e.g., finding an alternate route if a road is blocked)

• Utility-based Agents: These focus on optimality and long-term benefits, aiming for the "best" expected outcome (e.g., reaching a station with minimum fuel)

• Learning Agents: These are central to modern Machine Learning and Reinforcement Learning systems, allowing the agent to improve its performance over time

5. State Representations

The way an agent represents the world can vary in expressivity

• Atomic: The state is an indivisible "black box" (e.g., a specific chess board configuration)

• Factored: The state is split into various attributes like GPS coordinates or fuel levels

• Structured: Entities within the state have defined relationships (e.g., a traffic agent understanding the relationship between vehicles, pedestrians, and speed breakers)