3. Task Environments and the PEAS Framework

Source: AIMA 4th Ed, Chapter 2 (Section 2.3.1), physical PDF pp. 107–111

Introduction

Before building a rational agent, you must fully specify the problem the agent is meant to solve. Russell and Norvig call this the task environment — the formal setting in which the agent operates. The PEAS framework is the structured method for doing this specification.

The core idea: task environments are the “problems” to which rational agents are the “solutions.”

The PEAS Framework

PEAS stands for:

Letter	Component	What it specifies
P	Performance measure	How is agent success evaluated?
E	Environment	What does the agent interact with?
A	Actuators	How can the agent act?
S	Sensors	What can the agent perceive?

Design principle: In designing an agent, the first step must always be to specify the task environment as fully as possible using PEAS. Rushing to the algorithm before specifying PEAS is a common source of poorly designed AI systems.

Worked Example: Automated Taxi Driver (Figure 2.4)

This is AIMA’s primary extended PEAS example. It illustrates how a deceptively simple-sounding task resolves into a rich, multi-dimensional environment.

Performance Measure

The taxi should: - Get to the correct destination - Minimize fuel consumption and wear and tear - Minimize trip time or cost - Minimize traffic law violations and disturbances to other drivers - Maximize safety and passenger comfort - Maximize profits

Note that some of these goals conflict (speed vs. safety, profit vs. comfort), so the performance measure must encode tradeoffs. This is already non-trivial to specify correctly — the King Midas problem applies.

Environment

The taxi operates in: - Roads of all types: rural lanes, urban alleys, 12-lane freeways - Other traffic: cars, trucks, motorcycles, cyclists - Pedestrians and stray animals - Road works, police cars, puddles, potholes - Passengers (potential and actual) - Possibly different countries (drive on left vs. right)

The more restricted the environment is defined to be, the easier the design problem becomes.

Actuators

Steering wheel (direction)
Accelerator (speed up)
Brake (slow down)
Signal / horn
Display screen or voice synthesizer (communicate with passengers)
Possibly: inter-vehicle communication

Sensors

Video cameras (multiple, for 360° vision)
Lidar and ultrasound (distance to obstacles)
Speedometer (speed)
GPS (location)
Accelerometer (forces on vehicle)
Engine/fuel/electrical system sensors
Touchscreen or voice input (passenger destination requests)
Microphones

Complete PEAS table (Figure 2.4):

Agent Type	Performance Measure	Environment	Actuators	Sensors
Taxi driver	Safe, fast, legal, comfortable trip; maximize profits; minimize impact on others	Roads, other traffic, police, pedestrians, customers, weather	Steering, accelerator, brake, signal, horn, display, speech	Cameras, radar, speedometer, GPS, engine sensors, accelerometer, microphones, touchscreen

Additional PEAS Examples (Figure 2.5)

Agent Type	Performance Measure	Environment	Actuators	Sensors
Medical diagnosis system	Healthy patient, reduced costs	Patient, hospital, staff	Display of questions, tests, diagnoses, treatments	Touchscreen/voice entry of symptoms and findings
Satellite image analysis	Correct categorization of objects, terrain	Orbiting satellite, downlink, weather	Display of scene categorization	High-resolution digital camera
Part-picking robot	Percentage of parts in correct bins	Conveyor belt with parts; bins	Jointed arm and hand	Camera, tactile and joint angle sensors
Refinery controller	Purity, yield, safety	Refinery, raw materials, operators	Valves, pumps, heaters, stirrers, displays	Temperature, pressure, flow, chemical sensors
Interactive English tutor	Student’s score on test	Set of students, testing agency	Display of exercises, feedback, speech	Keyboard entry, voice

Software Agents (Softbots)

Virtual task environments can be just as complex as physical ones. A software agent (also called a softbot) operating on auction or reselling websites deals with: - Millions of users - Billions of objects (many with real images) - Dynamic pricing, competing agents, adversarial users

The PEAS description of a softbot is no less rich than that of a robot.

The Environment Class Concept

A single environment instance is rarely sufficient for evaluating agent performance. Experiments are typically run over an environment class — a distribution of environments sharing the same PEAS structure but varying in specific values (e.g., different traffic patterns, different weather, different passenger destinations for the taxi).

Agent performance is measured as average performance over the environment class
This prevents overfitting to a single scenario
The aima.cs.berkeley.edu code repository provides environment simulators for exactly this purpose

Why PEAS Matters

Prevents premature optimization. Without knowing the performance measure, you cannot even define what “better” means.
Exposes hidden complexity. The taxi example shows that what sounds like a simple driving task involves dozens of sensors, conflicting objectives, and an adversarial multi-agent world.
Informs algorithm choice. The PEAS description directly determines which agent architecture and algorithm family is appropriate (this is the subject of Section 2.3.2 — environment properties).
Enables evaluation. The performance measure component of PEAS is what you use to compare agent variants in experiments.

Common Exam Questions on PEAS

“Given agent X, write out its PEAS description” — practice with medical diagnosis, chess, spam filtering, warehouse robot
“What performance measure would you use for Y?” — watch for Goodhart’s Law traps (measuring the act vs. the desired state)
“How does the PEAS description change if you restrict the environment?” — e.g., taxi only in one city with perfect maps vs. globally deployed

Cross-References

Section 2.3.2 → Environment properties (the 6 dimensions — fully observable, deterministic, etc.)
Section 2.4 → Agent architecture selection is driven by the PEAS description
Chapter 17 → MDPs formalize the environment component of PEAS as a transition model
Chapter 22 → RL agents learn the performance measure from reward signals (the P component)