Anda di halaman 1dari 29

Arti%icial

Intelligence and Decision Systems


Inteligncia Ar.cial e Sistemas de Deciso MEEC, Maer,

1st Semester 2014/2015

Lus Manuel Marques Custdio



North tower ISR / IST

lmmc@isr.ist.utl.pt

Lus Custdio, Rodrigo Ventura

Course Goal and Syllabus


Provide background on basic no.ons and techniques used in Ar.cial
Intelligence, including:

Problem solving

Knowledge representa.on
Planning
Uncertainty
Decision making

Course Syllabus
Introduc)on: fundamental no.ons, historical aspects

Intelligent agents: percep.on/ac.on mapping, intelligent agent deni.on,
structure of agents, environments

Problem-solving: uninformed search, informed search, heuris.c func.ons,
adversarial search, constraint sa.sfac.on problems

Logical agents: knowledge representa.on and reasoning, proposi.onal logic, rst
order logic, inference by natural deduc.on, inference by resolu.on

Planning: planning agents, classical planning, Situa.onal calculus

Uncertainty: probabilis.c reasoning, Bayesian networks

Decision making: u.lity func.ons, decision networks, Markov decision processes

Distributed AI: mul.-agent systems (?)

Bibliography
Ar.cial Intelligence a Modern Approach

Stuart Russell and Peter Norvig

1st edi.on
avoid

2nd edi.on
okay

3rd edi.on
preferred

Student Evaluation
Wri`en evalua.on (70%)
2 tests: November 15, and December 20, or
Recourse exam: January 17
Prac.cal evalua.on (30%)
3 assignments: October 16, November 6, and December 4
these assignments involve solving three typical AI problems using
techniques learnt and Python

groups of 2 students

group registra.on in fenix

10 (9.5) mark minimum for approval, condi)oned of having at least 9.0


on each part

Class schedule
Theore.cal classes (28)

Tuesdays, 11h00-12h30, room EA5
Thursdays, 09h30-11h00, room EA4


Prac.cal classes (14)

Thursdays, 11h00-12h30, room V1.26
(rst 4 classes will be an introduc.on to Python)


Q&A (formal schedule)
Thursdays, 12h30-14h00, room 5.15
(send an email telling that you intend to go to Q&A)


Web page (at fenix):
h`ps://fenix.tecnico.ulisboa.pt/disciplinas/IASD25/2014-2015/1-semestre

Introduction
Ar)cial Intelligence is one of the newest elds in science and engineering: the
work started soon aker WWII
AI is a natural consequence of a .meless objec.ve:
Learning more about ourselves

What human characteris.c do you appreciate more?

In the past, physical strength was much more important than other human
characteris.cs, but now
(intelligence) knowledge is the key for success

Introduction
Ar)cial: made or produced by human beings rather than occurring naturally,
typically as a copy of something natural: her skin glowed in the ar.cial light | an
ar.cial limb | ar.cial owers.
Intelligence: the ability to acquire and apply knowledge and skills: an eminent
man of great intelligence | they underes.mated her intelligence.
[source: New Oxford American Dic.onary]

A very general mental capability that, among other things, involves the ability
to reason, plan, solve problems, think abstractly, comprehend complex ideas,
learn quickly and learn from experience. It is not merely book learning, a narrow
academic skill, or test-taking smarts. Rather, it reects a broader and deeper
capability for comprehending our surroundings"catching on," "making sense"
of things, or "guring out" what to do.
From "Mainstream Science on Intelligence" (Wall Street Journal 1994): a public statement
issued by a group of academic researchers in elds associated with intelligence

Introduction
Some concepts related with intelligence
Mind
Thought
Cogni.on

Percep.on

Knowledge
(representa.on)

Emo.ons

Reasoning

Behaviour

Learning
(memory)

Ac.on
9

and Evolu.on

Introduction
Derek Stubbs, a physician-turned-computer scien.st, claims that evolu.on of life
has made several great leaps, when judged by the criterion of adaptability:

Ability to reproduce
Sexual reproduc.on
Mul.cellular organisms
Development of specialized nerve cells
Inven.on of central nervous system
Ar.cial learning machines
Gene.c engineering, environment engineering and life-computer symbiosis

According to Hans Moravec, pioneer in mobile robot research and founder of


Carnegie Mellon University's Robo.cs Ins.tute, our robot crea)ons are evolving
similar to how life on Earth evolved, only at warp speed.

10

Introduction
We are very close to a .me where virtually no
human essen.al func.on, physical or mental, will
lack an ar.cial human emulator. The realiza.on of
this convergence of cultural developments will be
the intelligent robot, a machine that can think and
act like humans.
[Hans Moravec, 1988]

Here is a summary of Moravecs predic.ons for the future of robo.cs:


2010: A rst genera.on of broadly-capable "universal robots" will emerge. The servant robots will
be able to run applica.on programs for many simple chores.
2015: U.lity robots host programs for several tasks. Larger "U.lity Robots" with manipulator arms
able to run several dierent programs to perform dierent tasks may follow single-purpose home
robots.
2020: Universal robots host programs for most simple chores. Larger machines with manipulator
arms and the ability to perform several dierent tasks may follow, culmina.ng eventually in human-
scale "universal" robots that can run applica.on programs for most simple chores.
2030: Robot competence will become comparable to larger mammals.
11

RoboCup Federa)on goal: "By the middle of the 21st century, a team of fully autonomous humanoid
robot soccer players shall win a soccer game, complying with the ocial rules of FIFA, against the winner
of the most recent World Cup.

Introduction
Some ques.ons about human mind:

What is mind?
How can mental events occur in a physical world?
Do mental events arise from the physical, or are they some kind of spiritual
stu of their own?
If they do arise from the physical, how do they do it?
and if they dont, where do they come from?
Most researchers on AI and Robo.cs follow the physicalist assump.on:

Mind is what brain does, or something very like it in relevant ways

"If we really understand a system we will be able to build it. Conversely, we can
be sure that we do not fully understand the system un.l we have synthesized
and demonstrated a working model"
[Carver Mead, 1989]

That's the engineering point of view

12

Introduction
Approaches to mind

top-down
Ar.cial Intelligence

Psychology

Analy.c
perspec.ve

Mind

Neuroscience

Synthe.c
perspec.ve

Mechanisms of Mind
(Robo.cs)
bo`om-up

13

Introduction
Some examples of synthe.c approaches to mind:

Classical Symbolic AI
Ar.cial Neural Networks
Silicon Nervous Systems
Ar.cial Life
Computa.onal Neuroethology
Subsump.on Architecture (R. Brooks)
Society of agents (M. Minsky)
Most of these approaches are based on the following assump.ons:

Mind is be`er viewed as a con.nuous as opposed to a boolean no.on
Mind is aggregate rather than monolithic, i.e. it is enabled by a mul.tude of
disparate mechanisms
The main func.on of mind is to produce the next ac.on
Mind operates on sensa.ons to create informa.on for its own use
Mind uses prior knowledge (memories) to produce ac.ons by a reconstruc.ve
process, rather than by retrieval
Mind, to some degree, is implementable on machines

14

Introduction
Other interes.ng and related issues:

Animal mind
Rela.on between mind and body

Antnio Damsio

Book: Descartes' Error: Emo.on, Reason, and the Human


Brain, Putnam, 1994; revised Penguin edi.on, 2005

In order to re-center discussion back to AI, lets assume that Ar)cial Intelligence
is a synthe)c approach to mind

So, what dieren.ates AI from Philosophy or Psychology is that

besides trying to understand how the mind works,


AI also tries to design and develop ar)cial
en))es that somehow reveal to have
intelligence

[Charles Addams, The New Yorker, 2/9/1946. Tee and Charles Addams Founda.on.]

15

Introduction
The scien.c area of Ar.cial Intelligence was ocially
founded during the Dartmouth summer research project
on Ar.cial Intelligence (1956), involving John McCarthy,
Marvin Minsky, Nathaniel Rochester and Claude
Shannon.

The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature
of intelligence can in principle be so precisely described that a machine can be made to simulate it.


An a:empt will be made to nd how to make machines use language, form abstrac?ons and concepts,
solve kinds of problems now reserved for humans, and improve themselves.

Automa.c Computers
How Can a Computer be Programmed to Use a Language
Neuron Nets
Theory of the Size of a Calcula.on
Self-lmprovement
Abstrac.ons
Randomness and Crea.vity

16

Introduction
There are many deni.ons of Ar.cial Intelligence:
[the automa.on of] ac.vi.es that we associate
with human thinking



[Bellman, 1978]
The art of crea.ng machines that perform
func.ons that require intelligence when
performed by people



[Kurzweil, 1990]

The study of mental facul.es through


the use of computa.onal models

[Charniak e McDermo`, 1985]
The branch of computer science that is
concerned with the automa.on of
intelligent behavior

[Luger e Stubbleeld, 1993]

reasoning

think like humans

think ra.onally

behavior

act like humans

act ra.onally

human as model

ra.onal model

17

Introduction
What is ra.onality?

It is a kind of ideal performance:

doing the right thing given what it knows

Three types of ra.onality:



goals
ac)ons
beliefs
Isaac Asimov
(1920-1992)

Three laws of Robo)cs by Isaac Asimov



First Law: A robot may not injure a human being, or, through inac.on,
allow a human being to come to harm.
Second Law: A robot must obey the orders given it by human beings
except where such orders would conict with the First Law.
Third Law: A robot must protect its own existence as long as such
protec.on does not conict with the First or Second Law.

18

Introduction
Ac.ng humanly
Turing test

[Alan Turing, Computer machinery and intelligence, Mind, 59, 1950]

Computer needs:
Natural language processing
Knowledge representa.on
Automated reasoning
Machine learning
Illustra.on by Ann Witbrock in Copeland, B.J., Ar.cial Intelligence, Blackwell,Oxford, 1993

19

Does the human body serves no purpose, being only a mere sensor, even an
highly complex one?

Introduction
Ac.ng humanly
ELIZA
[Joseph Weizenbaum, mid 60s]

20

Introduction
Ac.ng humanly
Cleverbot

(h`p://www.cleverbot.com)

[Rollo Carpenter, 1997]

It is a web applica.on that uses an ar.cial intelligence algorithm to have


conversa.ons with humans.

21

Introduction
Thinking humanly

we must have some way of determining how humans think

by introspec.on
by psychological experiments
by brain imaging
Cogni)ve Modeling approach

more concerned with replica.ng human thought processes rather than
op.mizing performance
Cogni)ve Science

brings together computer models from AI and experimental techniques
from psychology to construct testable theories of the human mind

22

Introduction
Thinking humanly

Examples:

General Problem Solver (GPS), Alan Newell and Herbert Simon
(1959)
Neural Networks, McCulloch and Pi`s (1943)

Networks of McCullochPi`s neural elements.

Cogni.ve architectures: ACT-R (1983), SOAR (1987), etc.


23

Introduction
Thinking ra.onally

from laws of thought (philosophy) to mathema.cal logic

laws that govern the correct reasoning


Example:

Socrates is a man

all men are mortal


therefore Socrates is mortal
John McCarthy, Programs with common sense (1959)
This paper will discuss programs to manipulate in a suitable formal language (most likely

a part of the predicate calculus) common instrumental statements. The basic program will
draw immediate conclusions from a list of premises.

diculty on coping with: informal knowledge, uncertainty,


contradictory informa.on, bounded computer resources, etc.

24

Introduction
Ac.ng ra.onally

act in order to achieve the goals set, using the right means, given
current beliefs
Idea: develop ra.onal agents, en..es that interact with the environment
(perceive and act) using correct reasoning whenever possible

An agent is something that acts (from La.n agere, to do)


A ra)onal agent is one that acts so as to achieve the best outcome


in this course we will deal mainly with the general principles of ra.onal
agents and on components for construc.ng them

two notes on ra.onal agents:
if there is uncertainty, best outcome becomes best expected
outcome
if there is no sucient computer resources or .me, best outcome
is the best the agent can do with the available .me and resources
(limited ra)onality)

25

Introduction
Philosophy there are innumerous philosophical currents concerned with
issues related to thought

Socrates e Plato world of ideas vs. physical world

Aristotle
human reasoning

Descartes
dualism vs. materialism

Empiricists
knowledge comes from senses

Kant

hybrid posi.on between ra.onalism and empiricism

B. Russell
logical posi.vism

W. Quine
language is the key
26/26
Model of a robot knight based on drawings by Leonardo da Vinci.

Introduction

Mathema.cs many were the mathema.cians who studied and formalized


dierent aspects discussed by philosophers

Boole, Frege e Tarski logic as a model for formal reasoning

Hilbert e Godel
formal logical theories of natural numbers

(Godels incompleteness theorem)

Turing
computability

Karp

intractability
Psychology

classical


behaviorism


cogni.ve psychology

introspec.on
study of behavior
brain as an informa.on-processing device

27/26

Introduction
Ar.cial Intelligence

McCulloch e Pi`s

John McCarthy


Feigenbaum e Buchanan

Loki Zadeh

AI ages:

1943 55:
1956:
1952 69:
1966 73 :
1969 79 :
1980 ? :

neural networks
logic, LISP
expert systems
approximate reasoning (Fuzzy Logic)

gesta.on
birth
big hope
dark age
reborn
expansion

28/26

Introduction
AI was born following a discussion around the following ideas:

Some neuronal ac.vity consists of informa.on processing













[W. McCulloch, 1943]

A computer is mainly a symbolic manipulator













[C. Shannon, 1949]

Main research areas in AI



Symbolic processing (logic)
Connec.onism (neural networks)
Ar.cial Life (gene.c algorithms)
Reac.ve Robo.cs (bo`om-up approach)
Distributed AI, Mul.-Agent systems, Coopera.ve Robo.cs

29/26