ICS 481 Artificial Intelligence

1
ICS 481 - Artificial Intelligence

• Dr. Ken Cosh
• Lecture 1

2
This Weeks Topic

• What is A.I.?
• The History of Artificial Intelligence

3
A.I.

• Artificial Intelligence
• Artificial Made by humans produced rather
  than natural
• Intelligence The capacity to acquire and apply
  knowledge.
• courtesy Dictionary.com
• But Artificial Intelligence is more than just
  these 2 words put together, indeed - could
  artificial intelligence be considered an
  oxymoron?

4
Homo Sapiens

• Man the wise
• For years we have tried to understand how we
  think
• how does a bunch of stuff perceive, understand,
  predict and manipulate an environment more
  complicated than itself?
• A.I. is a new science which goes further
• as well as understanding how we do it, can we
  build intelligent entities?
• Exciting stuff!

5
So what is an A.I. system?

• 4 very different approaches
• One that thinks like humans
• One that thinks rationally
• One that acts like humans
• One that acts rationally

6
Different Approaches

• Human-centred vs Rational
• Is A.I. a science where we try to mimic human
  intelligence? Or do we combine maths and
  engineering to create rational intelligence?
• Thinking vs Acting
• Are we concerned with problem solving or decision
  making? Or are concerned with robotics or
  functionality?

7
Acting Humanly The Turing Test

• Arguably the most famous definition of
  intelligence
• Proposed by Alan Turing in 1950.
• Based on the premise that an entity is
  intelligent if it is indistinguishable from an
  undeniably intelligent entity - humans.
• A computer would pass the Turing test if an
  interrogator, having asked questions, cant tell
  whether the responses came from a computer or a
  human.
• The Turing test therefore takes a human-centred
  approach rather than a rational approach.

8
Passing the Turing test.

• To pass the Turing test, a computer needs to be
  able to mimic human responses.
• Consider asking a computer 1243225656423234462345
  3, compared to asking your friend?
• Consider asking your friend to tell you a joke,
  compared with asking a computer?

9
Passing the Turing Test

• To pass the Turing test a computer would need
  capabilities in the following
• Natural Language Processing
• Knowledge Representation
• Automated Reasoning
• Machine Learning
• To pass the Total Turing test (including physical
  simulation) add
• Computer Vision
• Robotics

10
Abstraction

• The quest for artificial flight succeeded when we
  stopped imitating birds and studied aerodynamics.
  Aeronautical engineering books do not define
  their goals as making machines that fly so
  exactly like pigeons that pigeons would be
  fooled! (Russell)
• Thus, the 6 disciplines highlighted on the
  previous slide compose much of what A.I. science
  investigates. Rather than duplicating an
  example, its more important to examine the
  underlying principles.

11
Thinking Humanly Cognitive Science

• To create a program which thinks like a human, we
  need better understanding of how the mind works.
• Introspection Catching our own thoughts as we
  have them.
• Psychological experiments Catching other
  peoples responses
• Once we have an accurate model of inputs,
  outputs, response timings etc. we can then create
  a program to match corresponding human behaviour.
• It is not enough to get the right answer, but you
  need to match a programs reasoning steps with the
  subjects.

12
Thinking Rationally

• Logic
• Given a correct set of premises, a program should
  always produce the correct conclusion
• Socrates is a man all men are mortal Socrates
  is mortal
• However, it isnt easy to take informal knowledge
  and state it formally in a logic notation - fuzzy
  logic.
• Secondly even a program with relatively few facts
  can require massive resources to process them -
  NP completeness.

13
Acting Rationally - Intelligent Agents

• An agent literally is something that acts, but
  computer agents are more than just programs.
• Autonomous Control
• Environment Perception
• Existing for a long time
• Adapting to change
• Taking on anothers goals
• A rational agent acts to achieve the best outcome.

14
Foundations of A.I.

• Here we examine some of the related fields of
  knowledge which have contributed to A.I.
• Philosophy
• Mathematics
• Economics
• Neuroscience
• Psychology
• Computer Engineering
• Control Theory Cybernetics
• Linguistics

15
Philosophy (428 B.C. to present)

• Aristotle began work on formulating laws that
  govern the rational mind
• I need a covering a cloak is a covering I need
  a cloak.
• What I need I have to make I need a cloak I
  need to make a cloak.
• This highlights goal based analysis (or a
  regression planning system). Starting with high
  level goals, and working backwards to find
  actions that move towards achieving the goals.
  This is now vital to agent based A.I..

16
Philosophy (428 B.C. to present)

• Philosophers such as Leonardo da Vinci designed
  automated computation devices, following a
  precise set of rules to produce results - such as
  arithmetic rules.
• The empiricism movement placed the origin of
  knowledge within the senses The principle of
  induction suggests rules can be created by
  repeated exposure to associations between
  elements.
• Later Logical Positivism states that all
  knowledge can be charaterised by logical theories
  stemming from sensory inputs.

17
Philosophy (428 B.C. to present)

• The confirmation theory extended this to
  investigate how knowledge gained from experience.
• Descartes was a supporter of Dualism, which
  proposes that there is a part of the mind that
  exists outside of the normal physical laws of
  nature, and it is this that allows us to decide
  not to fall towards the earth like a stone.
• All these philosophical investigations into the
  mind have shaped the way A.I. is investigated
  today.

18
Mathematics (800 to present)

• While Philosophers set out the ground rules for
  intelligence, mathematics transformed the rules
  into science. 3 key contributions from
  mathematics are
• Logic
• Computation
• Probability

19
Mathematics (800 to present)

• Logic
• Essentially Boolean logic
• Computation
• The development of algorithms to solve
  non-trivial problems.
• Algorithms analysis has taught us some problems
  are intractable - that is the time to solve them
  grows exponentially with the size of the
  instance.
• The PNP question, and hence NP-completeness are
  still significant problems today relating to A.I.
• Probability
• Invaluable for dealing with incomplete theories
  or uncertain measurements

20
Economics (1776 to present)

• Economics has made contributions to A.I. from
  when Adam Smith first published his inquiry into
  the wealth of nations.
• A science of how individual agents can maximise
  their benefits and a groups benefits.
  Essentially how people make choices that lead to
  preferred outcomes (mathematically as utility).
• This naturally leads into Decision Theory,
  helping intelligent agents make decisions, and
  Game Theory, leading to rational agents acting
  randomly on occasion.

21
Neuroscience (1861 - present)

• The study of the nervous system - predominantly
  in the brain.
• Neuroscience eventually taught us that the brain
  is the seat of consciousness and later that the
  brain is comprised of nerve cells which map to
  controlling different parts of the body.
• Sadly we still dont really know how all this
  works, but a collection of simple cells lead to
  thought, actions and consciousness
• unless you believe in mysticism!

22
Neuroscience (1861 - present)

• A.I. maps the brain into the computer.
• Moores law predicts that by 2020 CPUs will have
  as many gates as the brain has neurons.
• Moores law says the number of transisters per
  square inch doubles every 1 - 1.5 years, while
  the human brain capacity doubles every 2-4
  million years.
• But brains can act simultaneously.

23
Psychology (1879 to present)

• How do humans and animals think and act?
• Finding answers to this key question through
  introspection or observation.
• One area of Psychology has developed into
  cognitive psychology, which in turn led to
  cognitive science where first the information
  processing function of the brain is modelled
  using a computer, and later the psychology of
  memory, language and logical thinking.

24
Computer Engineering (1940 to present)

• A.I. requires intelligence and an artifact - and
  often computers are chosen as the artifact.
• From programmable machines, to operational,
  electronic, programmable computers
• As well as hardware, software has been developed
  to implement algorithms - for A.I. Lisp is often
  cited.
• A.I. has actually contributed back to the
  software field - linked lists stem from AI work.

25
Control Theory Cybernetics (1948 to present)

• Control Theory investigates how an artifact can
  modify its behaviour in response to changes in
  the environment - for instance maintaining
  constant water flow despite surges.
• Modern control theory is used in robotics, with a
  goal of maximising an objective function over
  time - very similar to the goal of AI.

26
Linguistics (1957 to present)

• Computational Linguistics, or Natural Language
  Processing, as a field of study came into being
  about the same time as AI, attempting to enable
  machines to understand natural language.
• This challenge is still a major challenge
  affecting AI.

27
History of A.I.

• Gestation (1943-1955)
• Before AI was officially born in 1956, various
  works examined the potential of neural networks
  and machine learning.
• Birth (1956)
• The name artificial intelligence was proposed by
  McCarthy from Princeton at a conference involving
  the early dominators of AI, from MIT, CMU,
  Stanford, IBM etc.
• AI became a field in its own right as it
  objectives and methodologies were different from
  any of the existing fields weve discussed.

28
History of AI

• Early Enthusiasm, Great Expectation (1952-1969)
• The early years were full of modest successes.
• General Problem Solver (GPS) was built to mimic
  human problem solving approaches through subgoals
  and possible actions - it was used to solve
  problems the intellectual establishment thought
  impossible.
• At IBM, a program was written to play checkers at
  a strong amateur level - this disproved that
  machines could only do what they were told to as
  it quickly learned to play better than its
  creator.
• The LISP programming language was developed -
  providing a tool for high level development.

29
History of AI

• A dose of reality (1966-1973)
• Most early AI examples ran by testing all
  different possibilities and choosing an optimal
  one, this just wasnt scalable! (Intractability)
• Huge grants were awarded by the US National
  Research Council to develop automated translation
  of Russian, but the problem proved harder than
  expected
• The spirit is willing but the flesh is weak was
  mistranslated as The vodka is good, but the meat
  is rotten
• Support for AI projects was cut and the field
  became weaker.

30
History of AI

• Knowledge Based Systems The key to power?
  (1969-1979)
• Early AI focused on weak methods, which werent
  scalable - while searching for the most
  applicable action amongst a small set of possible
  actions was effective, it didnt scale up to
  larger problems.
• The focus then changed towards handling problems
  from within a much tighter constrained domain.
• This lead to Expert Systems, where AI methods are
  applied to other areas of human expertise, such
  as medical diagnosis.

31
History of AI

• AI becomes an industry (1980 to present)
• Around the early 80s AI projects became
  commercially successful, with many businesses
  introducing expert systems with varying degrees
  of success.
• The return of neural networks (1986 to present)
• Neural networks as an approach had been abandoned
  in the 1970s as it had proved itself not to be
  useful - a 2 input perceptron could not be
  trained to recognise when its 2 inputs were
  different.
• However, the approach came back in the late 80s
  with parallel distributed processing.

32
History of AI

• AI becomes a Science (1987 to present)
• Rather than using toy examples to demonstrate
  weak theories, AI became more rigorously
  scientific where hypotheses are subjected to
  rigorous empirical experiments.
• This revolution has brought about speech
  recognition and data mining .
• The emergence of intelligent agents (1995 to
  present)
• Well investigate the contribution of intelligent
  agents shortly.

33
AI now

• What can AI do today?
• Autonomous planning and scheduling (detecting,
  diagnosing and recovering from problems aboard
  NASA spaceships)
• Game Playing (deep blue)
• Autonomous control (the computer controlled
  minivan that drove itself 2850 miles across
  America, with a human steering just 2 of the
  time)
• Diagnosis (Medical Diagnosis based on
  probabilistic analysis)
• Logistics Planning (Dynamic Analysis and
  Replanning Tool - DART, used by the US military
  during gulf war.)
• Robotics (for microsurgery)
• Language Understanding and problem solving

34
Assignment 1

• How could Introspection - reporting on ones
  inner thoughts - be inaccurate? Could I be wrong
  about what Im thinking? Discuss.