PEAS Framework in AI

PEAS

• PEAS stands for Performance, Environment, Actuators, Sensors
• Used to define task environments
• Performance Measure: Evaluates the agent's performance
• Environment: Where the agent operates
• Actuators: How the agent interacts
• Sensors: How the agent perceives

Example PEAS - Part-Picking Robot

• Performance measure: Percentage of parts correctly placed
• Environment: Conveyor belt with parts and bins
• Actuators: Jointed arm and hand
• Sensors: Camera and joint angle sensors

Example PEAS - Interactive English Tutor

• Performance measure: Maximize student's test score
• Environment: Set of students
• Actuators: Screen displaying exercises, suggestions, and corrections
• Sensors: Keyboard

Environment Types

• Fully Observable vs Partially Observable: Fully observable if sensors get all information about the environment
• Single Agent vs Multiagent: Single agent operates alone, multiagent operates in a competitive or cooperative environment
  • Multiagent challenges: Communication and randomized behavior
• Deterministic vs Stochastic: Deterministic if the next state is fully determined by the current state and agent action, otherwise stochastic
• Episodic vs Sequential: Episodic has independent experiences, sequential actions affect the future
• Static vs Dynamic: Static environments don't change while the agent decides, dynamic environments constantly change
  • Semidynamic environments do not change, but agent's score changes over time

Environment Types (continued)

• Discrete vs Continuous: Applies to environment state, time, percepts, and actions
• Known vs Unknown: Refers to the agent's knowledge about the environment's laws

Hardest Environment

• Partially Observable, Multiagent, Stochastic, Sequential, Dynamic, Continuous, and Unknown is the most difficult type
• Taxi Driving is hard in all these aspects, except it's generally known

Agent Structure

• Agent = Architecture + Program
• An Agent Program implements the agent function
• Should be designed based on the environment

Agent Programs

• Table-Driven Agent: Stores every possible percept sequence with a corresponding action
• Problems: Table size, creation time, learning time, guidance for entries

Agent Types

• Four basic types in order of increasing generality:
  • Simple Reflex Agents
  • Model-Based Reflex Agents
  • Goal-Based Agents
  • Utility-Based Agents
  • Learning Agents
  • All can be turned into learning agents

Simple Reflex Agents

• Select actions based on the current percept, ignoring history
• Examples: Vacuum cleaner agent
• Use condition-action rules: If car in front is braking, initiate braking

Problems with Simple Reflex Agents

• Only work if the correct decision can be made based on current percept
• Infinite loops are common in partially observable environments
• Randomization can help escape infinite loops

Model-Based Reflex Agents

• Maintain an internal state to represent unseen parts of the environment
• Use knowledge about how the world evolves and how actions affect the world
• Example: Driving requires tracking other cars that may not be in view

Goal-Based Agents

• Have a goal information defining desirable situations
• Example: Taxi needs to take in to account where it is trying to go
• Goal-based decision making can be straightforward or complex

Problems with Goal-Based Agents

• Less efficient than reflex agents, but more flexible
• Knowledge is explicitly represented and can be modified

Utility-Based Agents

• Have a utility function that measures how happy a state would make the agent
• Allows for comparison of different world states
• Example: Taxi could take multiple routes to its destination
• Rational decisions: utility function balances conflicting goals and likelihood of success against importance of goals

Learning Agents

• Can improve performance through learning
• Can operate in unknown environments
• Have a learning element to adjust performance based on feedback
• Learning element uses critic feedback and decides how to modify the performance component

Summary (continued)

• Agents interact with environments through actuators and sensors
• The agent function defines the agent's behavior
• The performance measure evaluates the agent's actions
• A rational agent maximizes expected performance
• Several agent architectures: reflex, reflex with state, goal-based, utility-based.