Artificial Intelligence

CS 404

Berrin Yanikoglu

Last update: Spring 2020

1
 Overview
• Course Info
– Briefly discussed. All of the information regarding the course
can be found at SUCourse.

• Introduction to AI

• AIMA – Chapter 1: Agents

2
• Instructor: Berrin Yanıkoğlu Office: MDBF 2056
• Office Hours: Fridays 9-10am or by email appointment.

• Textbook: Artificial Intelligence: A Modern Approach (AIMA), by Stuart Russell and

Peter Norvig., 3rd edition. I will give homeworks from the 3rd edition.

• Assistant:
– Furkan Coskun ([email protected]) and
– Ekberdjan Derman ([email protected])

• Important:
– Questions about homeworks etc should be asked at SUCourse discussion
boards. The discussion board will be followed every night at least, so you will get
an answer to your questions in about a day (often much sooner).
– Please dont ask homework related questions on email, so that the TA’s answer are
visible to everyone.
3
• Exams:
– There will be one midterm.
– Those who miss the midterm or the final exam needs to have a
documented excuse (medical emergency, death in immediate family) and
advance warning.
– In case of a missing exam (midterm or final), the makeup will be
counted for the missing exam. But if the makeup is taken due to non-
passing condition (low final or low overall score), the average of the final
and makeup exams will be used in computing the overall score.

• Homeworks:
– There will be 6 homeworks total, given approximately every other week.

• Grading:
– Midterm (35%) + Final (35%) + Homeworks (30%)
– To pass the course you grade as calculated above must be at least 40
(strict) and Final grade should be 30/100 or above.
4
Course Book & What we will cover

5
 Artificial Intelligence: A Modern
Approach Russel and Norvig
Part I Artificial Intelligence Part IV Uncertain Knowledge and Reasoning
1 Introduction 13 Quantifying Uncertainty
2 Intelligent Agents 14 Probabilistic Reasoning
15 Probabilistic Reasoning over Time
Part II Problem Solving
16 Making Simple Decisions
3 Solving Problems by Searching
17 Making Complex Decisions
4 Beyond Classical Search
5 Adversarial Search Part V Learning
6 Constraint Satisfaction Problems 18 Learning from Examples
19 Knowledge in Learning
Part III Knowledge and Reasoning
20 Learning Probabilistic Models
7 Logical Agents
21 Reinforcement Learning
8 First-Order Logic
9 Inference in First-Order Logic Part VII Communicating, Perceiving, and Acting
10 Classical Planning 22 Natural Language Processing
11 Planning and Acting in Real World 23 Natural Language for Communication
12 Knowledge Representation 24 Perception
6
25 Robotics
What is AI?

7
On May 12th, 1997, the best
chess player in the world, Gary
Kasparov, lost a six-game chess
match to a computer named
“Deep Blue 2”

• What was so significant

about this event?
Being able to program a computer to defeat a Grand Master level
chess player had been a long-standing goal of the science of
Artificial Intelligence and now it has been achieved.
 What is Artificial Intelligence?
Intelligence is difficult to define and understand,
even for philosophers and psychologists who spend
their lives studying it. But this elusive quality is, to
many people, the characteristic that sets humans
apart from other species

“What is intelligence, anyway? It is only a word that people use to name

those unknown processes with which our brains solve problems we call
hard. But whenever you learn a skill yourself, you are less impressed or
mystified when other people do the same.
This is why the meaning of “intelligence” seems so elusive: It describes not
some definite thing but only the momentary horizon of our ignorance
about how minds might work.”
- Marvin Minsky, AI researcher
 What is Artificial Intelligence?
Smart programs?

– Not really. Studying what is possible and underlying theories

are very important.

– How does a slow, tiny brain (biological or electrical) perceives,

understands, and manipulates a complex world?

10
 Studying AI
• For thousands of years, philosophers tried to understand
how humans think
• The field of Artificial Intelligence goes further and attempts
not just to understand but also to build intelligent systems
– Started out in 1950s
– The Dartmouth meeting in 1956
• Turned out much more difficult than anyone had imagined
• Currently encompasses a large variety of subfields,
– from general areas such as perception and logical reasoning to
– specific tasks such as playing chess, writing poetry…
– bringing together philosophy, logic, computer science, cognitive
science and cognitive neuroscience
11
 What is AI?
So it’s not surprising that defining artificial intelligence (AI) is hard. In general,
artificial intelligence is the field of science devoted to making computers
perceive, reason, and act in ways that have, until now, been reserved for human
beings.

4 types of historical definitions: acting or thinking humanly

acting or thinking rationally

12
 Thinking Humanly: Cognitive Science
• In order to say that a given program thinks like a human, we must have
some way of determining how humans thinks
• Requires scientific theories of internal activities of the brain
• What level of abstraction? ``Knowledge'' or ``circuits''?

• How to validate?
– Bring together computational models from AI and experimental
techniques from psycho-physics to model the human mind
• 1) Predicting and testing behavior of human subjects (Cognitive Science; top-
down)
• 2) Direct identification from neurological data (Cognitive Neuroscience;
bottom-up)
– Both approaches are now distinct from AI

– Most of the machinery of the human mind is not available to our

conscious experience
13
 Thinking Logically
• Let’s give up imitating the human thinking process, but try to
achieve logical thinking

• Laws of formal logic to formalize the thinking process

– Socrates is a man; All men are mortal => Socrates is mortal

• Difficulties:
– How to take informal knowledge and state in formal terms especially
when knowledge is less than 100% certain

– Even problems with just a few dozen facts can be prohibitive

14
Acting Humanly - Turing Test

15
• The computer passes the test if a human interrogator,
after posing some written questions, cannot tell
whether the written responses come from a person or
not
• Suggested major components of AI: natural language
processing, knowledge representation, automated
reasoning, machine learning
• Total Turing test also requires computer vision and
robotics
• “Artificial flight” is succeeded by not imitating the birds,
but by learning aerodynamics. The goal is not to fool
pigeons. 16
 Acting Humanly - Turing Test
• Reading: https://crl.ucsd.edu/~saygin/papers/MMTT.pdf

• Captchas
– Tests to identify humans from bots on the Internet, to deny services to
webcrawlers or spammers

17
 Acting Humanly - Turing Test
• Chatterbots:
– One of the most complex (and entertaining!)
chatter bots are at: http://www.simonlaven.com/

• choose complex chatterbots, choose Barry, go to

downloads and then choose CGI talk (this year they
dont have the online CGI talk, and the downloaded
program didnt work for me)

– Or this one works with no downloads: The ALICE

Foundation
• ...
• Human: You sound pretty smart
• ALICE: In what way do I sound like that?
Human: never mind, hey, how is the weather?
ALICE: Oh well thanks for trying to explain it to me anyway. You've
got my full attention. Sunny.
Human: where are you?
ALICE: I am in Dr. Richard S. Wallace's computer in Oakland,
California. Where are you?
Human: i am in istanbul, do you know where it is?
ALICE: Tell me a little about your life in istanbul do me know where
it is.
• ...

18
 Acting Rationally
• Rationality: ideal concept of intelligence

• Rational behavior: doing the right thing

• Thr right thing: what is expected to maximize goal achievement,
given the available information
– Reflexes do not necessarily involve thinking---e.g., blinking reflex,
but they are rational
– Thinking should be in the service of rational action

• Caveat: computational limitations make perfect rationality

unachievable
Þdesign best program for given machine resources

• We will emphasize rational agents in this course.

19
 Why do we want artificial intelligence?

• To relieve our mental labour, just as machines relieved

our physical labour last century

• It should make the machines themselves easier to use

• It might give some insight into the workings of our own

minds

20
AI Components

21
 Intelligent Agents
Machine Speech
Learning Understanding

Knowledge
Bilgi Gösterimi
Representation Computer
Search and Akıl Reasoning
Yürütme Vision
and
Optimization

AI
Natural
Language
Robotics Understanding

22
 Seeing, Hearing and Understanding
An intelligent computer must be able to recognize its surrounding environment
and adapt to changes in it. To do this it must be able to “see” and “hear”
what’s going on
Computer vision is the capability of a computer to mimic the ways that human
brains process and interpret light waves to produce a model of reality. Though
it’s very easy for people to do that, it’s very difficult for computers to do build
and update their models.
Computer Vision

From deepai.org
Hearing, Seeing and Understanding
The ability of a computer to recognize the speech of a user and
take action based on the words spoken is called speech
recognition or voice recognition. The computer matches spoken
words against stored speech patterns to determine what was said

Natural language processing is the ability of a computer to build

knowledge representations corresponding to the meaning in
sentences made up of recognized words. This is very difficult,
because human language is full of ambiguities, vagueness and
depends on a lot of commonsense knowledge of the world
Speech Understanding
 Natural Language Understanding
• Speech understanding: Sound to text
• Natural Language Understanding: Text to Meaning
• E.g.:“Evinin kapısında buluşalım”.
– Should we meet an your house door or his house
door?

27
 Natural Language Understanding
Natural language understanding (NLP) works well among languages
with similar structures (& maybe more importantly, for languages
for which there is enough data to learn from).
Had to correct an error in almost every other sentence in Turkish to
English translation (2019), with Google Translate:

28
 Machine Learning
•It is difficult to collect and maintain knowledge is.
•It would help if the machine could build up its own knowledge from
experiences in the world, like a child learning how to walk. The ability of the
machine to discover knowledge from observations of the world is called
machine learning.

For example, some of the best game-playing programs learn from past
experiences.
• If a move pays off, a learning program is more likely to use that (or
similar moves) in future games.
• If a move results in a loss, the program will remember to avoid
similar moves.
 Machine Learning
Supervised learning:
Learning from
examples

Kadın

Training Testing
 Knowledge Representation and
Reasoning
• Rain => WetGround
• Sprinkler => WetGround
• WetGround => Rain OR Sprinkler

31
 Robotics
• Japanese companies such
as Honda, Fujitsu and Sony
are racing to develop
humanoid robots.

• e.g. Honda ASIMO (right)

• Improved walking stability

over earlier models

• Smaller size is about

marketing – and Robocup
eligibility

• Intelligence quite limited –

some commands sent by
remote control

• Will cost about the same a

luxury car.
History of AI

33
 History of AI
• 1943 McCulloch and Pitts: Artificial Neuron Model
• 1950 Turing's ``Computing Machinery and Intelligence'‘

• 1950s Early AI programs, including Samuel's checkers program, Newell & Simon's
Logic Theorist (proving theorems), Gelernter's Geometry Engine, Shannon
and Turing writing chess programs
– Shortage of computer times => Development of time sharing (=> DEC)
– Creation of LISP (McCarthy)

• 1956 Dartmouth meeting: ``Artificial Intelligence'' coined

• 1965 Robinson's complete algorithm for logical reasoning resolution method

• 1960sEarly development of knowledge-based systems; Minsky’s microworlds (blocks

as home to various projects: vision, planning, nat. Lang. Understanding, ...)
– ANALOGY program (what is this figure most similar to?)
– Algebra STUDENT program (one egg costs ... How much does twenty eggs cost?)
34
 History of AI
• 1966--74 Dose of Reality

• Very little domain knowledge:

– Swithing from one domain to another, the programs failed miserably

• AI discovers computational complexity

– Early programs worked by representing the basic facts and trying out a
series of steps to solve the problem which was only tractable within
micro worlds; NP-completeness showed that scaling up to larger
problems was not always viable

• Neural network research almost disappears

35
 History of AI
• 1980--88 Expert systems industry booms
– After all, they work, even if in limited domains
– An expert system is a software designed to replicate the decision-making
process of a human expert, within a narrow topic. At the heart of every
expert system is a knowledge base representing ideas from the specific
field of expertise
– A knowledge-based system derives knowledge from experts as well as
other sources like government regulations, statistical databases, company
guidelines, etc.
– In practice, the terms expert system and knowledge-based system are
often used interchangeably
• While a database contains only facts, a knowledge base also contains a
system of if-then rules for determining and changing the relationships
between those facts

36
 History of AI
• 1988--93 Expert systems industry start losing its power
• Successful only in very narrow domains
• Building a successful expert system is much more than simply buying a
reasoning system and filling it with rules

• 1985--95 Neural networks return to popularity

• 1988-- With strengthened foundations, AI becomes hot again -

resurgence of probabilistic and decision-theoretic methods, genetic
algorithms, belief networks,...

• 2000 -... Deep learning, significant advances in robotics

Now we are talking about the perils of AI

38
Current State of AI

39
 Narrow AI

Google – autonomous vehicle 2019

 Goal: General AI
Human-like intelligence,not in only a narrow area:

Sophia - Hanson Robotics 2018

 Current State
• Which of the following can be done at present?

– Play a decent game of table tennis

– Drive along the highway
– Play a decent game of bridge
– Give competent legal advice in a specialized area of law
– Translate spoken English into spoken Swedish in real time
– Discover and prove a new mathematical theorem
– Write an intentionally funny story (relatively bad jokes)
– Paint/Compose along given style

42
 Current State
• Which of the following can be done at present?

– Play a decent game of table tennis

– Drive along the highway
• Drive in the center of Istanbul
– Play a decent game of bridge
– Give competent legal advice in a specialized area of law
• Converse on general topics
– Translate spoken English into spoken Swedish in real time
• Translate Turkish to English well in real time
– Discover and prove a new mathematical theorem
– Write an intentionally funny story
• Not very funny J
– Paint/Compose along given style
• Originality?
43
 Current State

• Limited domain speech/natural language understanding

programs

• Chess playing programs (machines)

• Medical expert systems challenging doctors

• ...

44
 Artificial Intelligence and the Humans

•What does the advent of the intelligent machine mean for human
beings?
•Are artificial intelligences just extensions of human intelligence?
•When AARON creates a drawing, who is the artist, Cohen or AARON?
•When expert systems make decisions, who is responsible? the user, the
programmer, the software company, or somebody else?

•Should we think of intelligent machines as some new sort of life, one

with which we must now share the world?

•Could AIs be our evolutionary successors?

•How will AI affect our own sense of self?

•AI is beginning to force us to confront these hard philosphical

questions…
 To Know
• Basic history of AI
– Know some of the important events or at least what
happend in different eras and current state
• Difficulty of defining intelligence and some of the
attempts
– Fleeting nature of the definition
– Difference of humanly/rational thinking/acting
• Turing test
• Components of an AI agent
• Narrow and general AI concepts
46
 Reading
• AIMA Chapter 1
• “Computing Machinery and Intelligence”, A. Turing

47