An agent perceives its environment through sensors and acts on it through actuators. Describing an AI problem starts by pinning down four things — its PEAS — and the properties of its environment, which decide which techniques will work.

Environment → Sensors → percept → Agent program → Actuators → action → Environment

The vacuum world

The textbook toy: two squares, A and B, each clean or dirty. The agent perceives (location, status) and can Suck, move Left or Right. The performance measure awards one point per clean square per time step:

Python

def reflex_vacuum_agent(percept):
    location, status = percept
    if status == "Dirty":
        return "Suck"
    return "Right" if location == "A" else "Left"

def run(agent, world, location="A", steps=6):
    score, actions = 0, []
    for _ in range(steps):
        action = agent((location, world[location]))
        actions.append(action)
        if action == "Suck":
            world[location] = "Clean"
        elif action in ("Left", "Right"):
            location = "B" if action == "Right" else "A"
        score += sum(s == "Clean" for s in world.values())   # performance measure
    return actions, score

actions, score = run(reflex_vacuum_agent, {"A": "Dirty", "B": "Dirty"})
print(actions)    # ['Suck', 'Right', 'Suck', 'Left', 'Right', 'Left']
print(score)      # 10 ← 1 + 1 + 2 + 2 + 2 + 2

PEAS — specifying the task

Agent	Performance measure	Environment	Actuators	Sensors
Self-driving taxi	Safe, fast, legal, comfortable trip; profit	Roads, traffic, pedestrians, weather, passengers	Steering, accelerator, brake, signals, display	Cameras, LiDAR, radar, GPS, speedometer, odometer
Medical diagnosis system	Healthy patient, low cost, few tests	Patient, hospital, staff	Display questions, tests, diagnoses, treatments	Keyboard entry of symptoms, test results
Chess program	Win / draw / loss	Board, opponent, clock	Moves on the board	Board state
Spam filter	Spam caught, legitimate mail kept	Mail stream, users	Mark as spam / deliver	Message text, headers, sender reputation

Properties of task environments

Dimension	Easy end	Hard end	What it changes
Observability	Fully observable — sensors show the whole relevant state	Partially observable — noisy sensors, hidden state	Hidden state → agent must keep internal state / beliefs
Agents	Single-agent	Multi-agent — competitive or cooperative	Opponents → game theory, adversarial search
Determinism	Deterministic — next state fixed by state + action	Stochastic (nondeterministic) — outcomes uncertain	Uncertainty → probabilities, expected utility
Episodes	Episodic — each decision independent (classify one image)	Sequential — actions affect the future (chess)	Sequential → planning ahead
Change	Static — world waits while agent thinks	Dynamic — world changes meanwhile	Dynamic → time pressure, anytime algorithms
State/time	Discrete — finite states and actions	Continuous — positions, speeds	Continuous → optimisation, control theory
Knowledge	Known — rules/physics given	Unknown — must learn how the world works	Unknown → exploration, learning

Task	Observable	Agents	Deterministic	Episodic	Static	Discrete
Crossword puzzle	Fully	Single	Yes	Sequential	Static	Discrete
Chess with a clock	Fully	Multi	Yes	Sequential	Semi	Discrete
Poker	Partially	Multi	Stochastic	Sequential	Static	Discrete
Taxi driving	Partially	Multi	Stochastic	Sequential	Dynamic	Continuous
Image classification	Fully	Single	Yes	Episodic	Static	Continuous

What does PEAS stand for?

Is a rational agent always successful?

What's the difference between the agent function and the agent program?