The concept of Agents and Environments - Artificial Intelligence

The agent is a central abstraction in AI. An agent:

Perceives its environment through sensors
Acts upon the environment through actuators
Makes decisions based on perceptions to achieve goals

An agent can be a robot, a chatbot, or any other entity that can perceive and act. The decisions of an agent are made in the context of its environment.
An environment is the surrounding or conditions in which an agent operates.

1. Formal Definition:¶

An agent function maps from percept histories to actions:

f: P^* \rightarrow A

(1)

where $(P^*)$ is the set of all possible percept sequences and (A) is the set of possible actions.

2. PEAS Description¶

AI agents are characterized using PEAS (Performance, Environment, Actuators, Sensors):

Agent Type	Performance	Environment	Actuators	Sensors
Cleaning Robot	Cleanliness, battery, safety	Room, furniture, dirt	Wheels, vacuum	Camera, dirt sensor
Chess Program	Win/loss/draw	Chess board, opponent	Move selection	Board position
Self-driving Car	Safety, speed, comfort	Roads, traffic	Steering, throttle	Cameras, LIDAR, GPS
Medical Diagnosis	Patient health	Patient symptoms	Treatment plan	Lab tests, imaging
Chatbot	User satisfaction	Conversation context	Text response	User input text

3. Types of Agents¶

3.1. Simple Reflex Agent¶

Acts based only on current percept, ignoring history.

Structure:

function REFLEX-AGENT(percept):
  if percept matches condition C1:
    return action A1
  if percept matches condition C2:
    return action A2
  ...

Example: Thermostat - if temperature < 20°C, turn on heater

Advantages: Simple, fast, minimal memory Limitations: Cannot handle partially observable environments, no memory of past

3.2. Model-Based Agent¶

Maintains an internal state to track aspects of the world.

Structure:

function MODEL-BASED-AGENT(percept):
  state ← UPDATE-STATE(state, action, percept, model)
  action ← CHOOSE-ACTION(state, rules)
  return action

Example: Vacuum cleaner remembering which rooms it has cleaned

Advantages: Handles partial observability, remembers history Limitations: More complex, requires accurate world model

3.3. Goal-Based Agent¶

Has explicit goal information and plans actions to achieve it.

Structure:

function GOAL-BASED-AGENT(percept):
  state ← UPDATE-STATE(state, action, percept, model)
  plan ← FIND-PLAN(state, goal)
  action ← FIRST-ACTION(plan)
  return action

Example: GPS navigation - planning route to destination

Advantages: Can plan ahead, flexible to new goals Limitations: Planning can be computationally expensive

3.4. Utility-Based Agent¶

Chooses actions to maximize expected utility (degree of happiness/success).

Structure:

function UTILITY-BASED-AGENT(percept):
  state ← UPDATE-STATE(state, action, percept, model)
  action ← argmax over actions of EXPECTED-UTILITY(action, state)
  return action

Example: Stock trading bot - maximizing profit while managing risk

Advantages: Can handle conflicting goals, trade-offs, uncertainty Limitations: Requires utility function, more computational overhead

3.5. Learning Agent¶

Improves performance through experience.

Components:

Performance Element: Selects actions
Learning Element: Improves performance element
Critic: Provides feedback on performance
Problem Generator: Suggests exploratory actions

Example: AlphaGo - learned to play Go through self-play

Advantages: Adapts to new situations, improves over time Limitations: Requires training data/experience, may converge slowly

4. Types of Environments¶

Fully observable: The agent has access to all relevant information about the environment at all times.
Partially observable: The agent has limited access to information.
Deterministic: The outcome of an action is predictable and certain.
Stochastic: The outcome of an action is uncertain and may involve randomness.
Static: The environment does not change while the agent is deliberating.
Dynamic: The environment can change while the agent is deliberating.
Discrete: The number of possible states and actions is finite.
Continuous: The number of possible states and actions is infinite.

The characteristics of an environment significantly affect agent design:

Dimension	Type 1	Type 2	Example
Observability	Fully observable	Partially observable	Chess vs. Poker
Determinism	Deterministic	Stochastic	Chess vs. Backgammon
Episodes	Episodic	Sequential	Image classification vs. Chess
Dynamics	Static	Dynamic	Crossword vs. Self-driving
State Space	Discrete	Continuous	Chess vs. Robot control
Agents	Single-agent	Multi-agent	Puzzle vs. Poker

Fully Observable: Agent can see complete state (e.g., chess board) Partially Observable: Agent sees only part of state (e.g., poker - can’t see opponent’s cards)

Deterministic: Next state completely determined by current state and action Stochastic: Randomness affects outcomes (e.g., dice, unpredictable opponents)

Episodic: Each episode is independent (e.g., classifying images) Sequential: Current decisions affect future (e.g., chess moves)

Static: Environment doesn’t change while agent deliberates Dynamic: Environment changes continuously (e.g., real-time strategy games)