Introduction

C0: Introduction

1

Most materials in this chapter are based on materials from https://ai.berkeley.edu

C0

Introduction

COMP3270

Introduction to Artificial Intelligence

Today we ask what artificial intelligence means. By the end, we will view AI as agents that perceive an environment and choose actions.

What Would Count as Intelligence?

What Would Count as Intelligence?

2

TASK

SOLUTION

ACalculator

NoExecutes fixed rules for a supplied calculation

BChess Engine

YesCompares possible moves and chooses one toward its goal

CChatbot

MaybeHuman-like responses do not necessarily prove understanding

DSelf-Driving Car

YesPerceives change and adjusts its actions

What evidence are you using?

Start with your intuition. Which systems count as intelligent, and what evidence supports your answer?

Reasonable answers can differ because the criterion differs. That disagreement gives us our first question: how should intelligence be judged?

How Should We Judge Intelligence?

How Should We Judge Intelligence?

3

What do we evaluate?

THINKING ↔ ACTING

What is the benchmark?

HUMAN-LIKE ↔ RATIONAL

Combine the answers → four ways to define AI

We can organize those criteria along two axes. First, do we evaluate internal thinking or observable action?

Second, do we compare the system with humans or with a rational standard?

Together, those choices produce four approaches to AI.

Build AI That…

Build AI That…

4

HUMANLY

RATIONALLY

THINK

Think HumanlyThinks like a human

Think RationallyReasons correctly

ACT

Act HumanlyActs like a human

Act RationallyChooses effective actions

Here are the four approaches in one map.

Thinking humanly asks whether a model explains how people actually think.

Acting humanly asks whether behaviour resembles human behaviour.

Thinking rationally asks whether conclusions follow valid rules.

Acting rationally asks whether actions achieve the objective. This is our course perspective.

Which Approach Guides Each AI System?

Which Approach Guides Each AI System?

5

TASK

SOLUTION

AI System

Approach

A hypothetical brain simulation that reproduces how humans think

Think Humanly

A chatbot designed to convince a judge that it is human

Act Humanly

An automated theorem prover that derives valid conclusions

Think Rationally

A self-driving car that chooses actions to reach its destination safely

Act Rationally

Apply the map. Decide which approach most directly guides each system.

The brain simulation tries to reproduce human thought, so it is thinking humanly.

The chatbot is judged by human-like behaviour, so it is acting humanly.

The theorem prover derives valid conclusions, so it is thinking rationally.

The car chooses actions toward its goal, so it is acting rationally.

Thinking Humanly Models Human Cognition

Thinking Humanly Models Human Cognition

6

TASK

SOLUTION

OBSERVE Human Thinking Introspection · experiments · brain activity

→

MODEL Computational Process Turn a theory into a program

→

COMPARE Model and People Do they solve the task in the same way?

A person and a model give the same answer. Does that mean they thought in the same way?

No. Matching answers show the same behaviour—not the process that produced it.

Thinking humanly starts with evidence about how people think.

We turn a theory of that process into a computational model.

Now compare the model with people. If both produce the same answer, did they necessarily use the same process?

No. Matching answers show the same output, not the process that produced it.

Acting Humanly Tests Human-Like Behaviour

Acting Humanly Tests Human-Like Behaviour

7

Can a judge reliably tell which response came from a person?

?

HUMAN JUDGE

↔ TEXT ONLY

HUMAN

AI

The test evaluates observable behaviour, not the hidden process

Acting humanly evaluates behaviour. Turing’s imitation game makes that idea operational.

A judge exchanges text with one human and one AI without seeing either.

If the judge cannot reliably tell them apart, the AI passes this behavioural test. It need not think like a human.

Thinking Rationally Applies Correct Reasoning

Thinking Rationally Applies Correct Reasoning

8

TASK

SOLUTION

All men are mortal

Socrates is a man

Therefore, Socrates is mortal

What is the limitation of this approach?

Many real-world problems cannot be fully captured by logical facts and rules. Knowledge may be informal, incomplete, or uncertain.

Thinking rationally uses logic: true premises and a valid rule produce a correct conclusion.

From these premises, Socrates must be mortal.

What limits this approach? Consider knowledge that is incomplete, uncertain, or difficult to formalize.

Many real problems cannot be captured completely by logical facts and rules. Logic is powerful, but not sufficient alone.

This Course Focuses on Acting Rationally

This Course Focuses on Acting Rationally

9

DEFINITION

Choose the action expected to achieve the best outcome

This course focuses on acting rationally: choosing the action expected to produce the best outcome. The word expected matters when outcomes are uncertain.

Example: Acting Rationally Under Uncertainty

Example: Acting Rationally Under Uncertainty

10

TASK

SOLUTION

FORECAST 70% chance of rain

→

ACTION ☂ take an umbrella

→

OUTCOME ☀ it stays dry

Was taking the umbrella irrational?

No. A good decision can still have an unlucky outcome. Judge the choice using what was known at the time.

A 70% chance of rain makes taking an umbrella reasonable. But suppose it stays dry: was the decision irrational?

No. A good decision can have an unlucky outcome. Judge the action using the information available at the time.

An Agent Perceives, Decides, and Acts

An Agent Perceives, Decides, and Acts

11

An agent is anything that perceives its environment and acts upon it.

The loop repeats as the world changes

An agent is anything that perceives its environment and acts upon it.

The environment sends percepts: the observations available to the agent.

The agent chooses actions that affect the environment, and the loop repeats. This course studies the decision inside that loop.

What Made This Choice Rational?

What Made This Choice Rational?

12

Taking the umbrella depended on four things

OBJECTIVEStay dry

EVIDENCE70% chance of rain

KNOWLEDGEAn umbrella reduces the risk

ACTIONSTake it or leave it

Why was taking the umbrella rational? Rationality depends on the situation, not on the action alone.

The objective was to stay dry.

The evidence was a 70% chance of rain.

The agent knew that an umbrella reduces the risk.

The available actions were to take it or leave it. Change any one of these four factors, and the rational choice may change.

Rational Does Not Mean Omniscient

Rational Does Not Mean Omniscient

13

RATIONAL ≠ KNOWS THE FUTURE

Use the information available now

Choose under limited time and memory

A rational decision can still lead to a bad outcome

Rational does not mean knowing the future.

A rational agent uses the information available now.

It also has limited time and memory. A rational decision can therefore still end badly.

A Self-Driving Car With One Objective

A Self-Driving Car With One Objective

14

TASK

SOLUTION

SELF-DRIVING CARMaximize safety

Which behaviour receives the highest score?

Never leave the garage

Suppose a self-driving car has one objective: maximize safety. Which behaviour earns the highest score?

Never leave the garage. It is perfectly safe but useless because the objective omitted progress.

A Vacuum Robot With One Objective

A Vacuum Robot With One Objective

15

TASK

SOLUTION

VACUUM-CLEANING ROBOTCollect as much dirt as possible

How could the robot maximize its score?

VACUUM DIRT

→←

DUMP IT

The score rises—even though the floor never gets cleaner.

Now reward a vacuum robot for the amount of dirt it collects. How might it maximize that score?

It can vacuum dirt, dump it, and collect it again. The score rises while the floor stays dirty.

Specify the Performance Measure Carefully

Specify the Performance Measure Carefully

16

PERFORMANCE MEASURE How the agent's success is scored

POOR MEASUREAmount of dirt collected

BETTER MEASURECleanliness maintained over time

ALSO CONSIDERcoverage · energy · time · safety

A performance measure is how we score success. It turns the intended goal into something operational.

Amount of dirt collected is poor because the same dirt can be counted repeatedly.

Cleanliness maintained over time better represents what we actually want.

A realistic measure may also balance coverage, energy, time, and safety. Choosing the measure is part of designing the agent.

This Course Builds a Toolkit for Rational Action

This Course Builds a Toolkit for Rational Action

17

SEARCH · C1Which action sequence reaches the goal?

MDPs · C2Which action is best when outcomes are uncertain?

REINFORCEMENT LEARNING · C3How can an agent improve from experience?

PROBABILISTIC MODELS · C4–C6What should the agent believe from noisy evidence?

LOGIC · C7What follows from what the agent knows?

This course develops five tools for rational action.

C1 uses search to find an action sequence that reaches a goal.

C2 uses Markov decision processes to choose actions when outcomes are uncertain.

C3 uses reinforcement learning to improve through experience.

C4 to C6 use probabilistic models to reason from uncertain evidence.

C7 uses logic to represent knowledge and draw valid conclusions. The right tool depends on what the agent knows and how the environment behaves.

AI Progress Comes in Waves

AI Progress Comes in Waves

18

1950sLogic and Search

1970s–80sExpert Systems

1990s–2000sProbability and Data

2010sDeep Learning

2020sGenerative and Agentic AI

Each wave solves some problems and reveals new limitations

These ideas did not arrive all at once. AI has progressed in waves.

Early AI emphasized logic, problem solving, and search.

Expert systems encoded specialized human knowledge.

Probability and data made uncertainty and learning central.

Deep learning transformed perception and representation learning.

Generative and agentic AI expanded the range of tasks again. Each wave solves some problems and exposes new ones.

The Thinking Machine, 1961

The Thinking Machine, 1961

19

TASK

What stands out about the way they describe intelligent machines?

SOLUTION

Their enthusiasm is striking: major breakthroughs feel close

We will end with a view from 1961. As you watch, listen to their confidence: what stands out about how they describe intelligent machines? Play the excerpt.

Their enthusiasm is striking: major breakthroughs feel close. Many of their ambitions, including search, learning, and language, still shape AI today.

Reference Materials

Reference Materials

20

Course Details 2-page course guide Course focus, schedule, assessment, and policies AIMA: Chapter 1 66 pages comp3270.dirk.hk/AIMA_Chapter_1.pdf  ·  Password: Schnieders AIMA: Chapter 2 49 pages comp3270.dirk.hk/AIMA_Chapter_2.pdf  ·  Password: Dirk The Thinking Machine: MIT 1961 53-minute film comp3270.dirk.hk/the_thinking_machine_mit_1961.mp4

The course-details PDF summarises the structure, schedule, assessment, and policies. The other materials let you explore the chapter in more depth. Each row opens privately in the Demo Browser.

From Rational Agents to Search

From Rational Agents to Search

21

An agent has a goal and several possible actions

How can it find a good action sequence without exploring every possibility?

Next, we turn the rational-agent perspective into an algorithm. An agent has a goal and several possible actions.

Each action creates new states and more possible action sequences.

The goal is somewhere in this growing tree. C1 asks how to find a good sequence without exploring every possibility.