Redman322presentation AI

Download Report

Transcript Redman322presentation AI

Artificial
Intelligence
Where it has been and where it
should be going
AI is conceived….(slowly)
 6th Century - First mention of any type of automata in Homer’s
Iliad, with the mention of Hephaestus’, the God of Fire, workshop.
 5th Century- Aristotle codifies the first formal deductive reasoning
system: Syllogistic logic.
 15th and 16th Century- Mechanical clocks first appear in Europe.
 1642- Pascal invents the mechanical calculator, the “Pascaline.”
 1673- Leibniz improves upon Pascal’s calculator and envisions a
“universal calculus of reasoning” to decide arguments
mechanically.
 19th Century- AI in literature: Hoffman’s The Sandman, Goethe’s
Faust, and Shelley’s Frankenstein.
AI is articulated… (sort of)
 Babbage produces a small working model of
the Difference Engine and persuaded the
British government to finance a larger model.
 Babbage couldn’t return what he promised.
 Babbage ignores the Difference Engine in
pursuit of the Analytical Engine, a machine that
could do not only arithmetical calculations but
also analysis and reasoning.
The AI story continues
 Ava Lovelace, Byron’s daughter, takes an
interest in Babbage’s work.
 1890- Herman Hollerith performs the US
census using a machine that encoded
information using punch cards.
 1923- The term “robot” is introduced into
the English language in Karel Capek’s
play, Rossum’s Universal Robots.
Turing
(this guy didn’t smile very often
so this pic was a find)
 Turing Test: Any machine that could hold a
conversation with a person who cannot see
who he is talking and cannot distinguish
whether it is a machine or human is intelligent.
 Universal Machine: A machine that has the
ability to learn any number of tasks and
perform them. (Humans are a universal
machine or at least close enough.)
 Intelligence comes from organization.
Cybernetics



Communication within both living being and machines
Norbert Wiener says this is the source of intelligent behavior.
This is obviously not a process limited to human beings and thus
intelligence is not a characteristic limited humans.
John Von Neumann
(no luck on a smile from this guy)
 1942- Von Neumann comes to Princeton convinced
that high-speed computing is the essential to the future
of science and mathematics, and sets to work on the
design of the IAS machine.
 The IAS’s predecessors, ENIAC and EDVAC, were
designed for specific tasks (nautical and bombing
tables). The IAS could be manipulated to solve a wide
class of problems if only they were input properly.
 1951- “anything that can be completely and
unambiguously described , anything that can be
unambiguously put into words is ipso facto realizable
by a suitable finite neural network.”
The Dartmouth Conference
(Dartmouth College… artists rendering)
 The term “artificial intelligence” is
adopted, to the chagrin of many of those
present
 McCarthy proposed the idea of a programming
language for use in problems requiring
conjecture and self-reference.
 “We are a long way from even knowing what
questions to ask or what aspects to abstract for
theory. The present need is for a large
population of concrete systems that are
completely understood…”
The InformationProcessing Model
 Redefine the computer as an information
processor not a calculating machine.
 symbolic-functioning capabilities of
computers.
 a symbol is something that could stand
for an object and its meaning, its uses,
and all other pertinent information about
it.
Games
kerplunk ->
<- obvious
 The games computers played at first were the ones that were
most interesting from a computational standpoint: chess and
checkers.
 Samuel considered a different approach, instead of imitating a
human thought process he wanted to enlist a new thought process
all together.
 “I think you study the way people solve problems to get an insight
into what the real problem is…And then you sit down and say,
‘Okay, given the technology available…how best can we solve the
problem?”
 “The machine, therefore, played a perfect ending without one
misstep. In the matter of the end game, I have not had such
competition from any human being since 1954, when I lost my last
game.”
The Game: Chess
 Other programs of the day took advantage of the speed of the
computer processor, using it to search the field of possible board
configurations after each of its various possible moves and then
choosing the optimal configuration.
 “if all you have is a machine that bests humans by means of
speed alone, what really do you have?”
 Economizing search strategies among a set of possibilities has
proven a key element in Artificial Intelligence to date.
 Although we cannot mimic bird flight, we can still fly anywhere we
want.
 Side Note: In 1997 IBM’s Deep Blue defeated Garry Kasparov, the
world’s reigning chess champion at that time.
Vs.
Pseudo-Intelligent
Systems
(Cyrus Whitney far cooler than
either of the two listed below)
 PARRY – Imitated a paranoid patient and was
made to be interviewed by psychiatrists
 PROSPECTOR – Analyzes geologic data and
makes interpretations and predictions.
 LUNAR – Contained a wealth of information
about moon rocks and was able to answer
questions about them written in plain English
 CYRUS – Was fed news articles about
Secretary of State Cyrus Vance and was able
to be interviewed as if it were him.
They All Fall Short
 While impressive these systems only
manage to simulate a small range of
thinking.
 None of them are able to provide any
meaning to the data they present.
 Still a long way from Turing’s Universal
machine.
<- Yao Ming
Ryan Foss->
Robotics
Isaac Asimov this
guy wrote a lot of
stories about
robots… but never
met one
 In the late 1960s three major robotics projects got started in the
United States. Each project had its own particular flavor but the
general idea was to create some sort of independent thinking,
interacting, machine.
 Robotics (a term coined by Isaac Asimov) was the practical
application of AI, and it ended up having a significant impact on
the theory.
 Requirements for AI were rewritten, and a system now had to be
able to demonstrate that it had an internal model of the world, be
clever enough to answer questions on a wide range of topics
(analytical and common), acquire information from the external
world and update its model, and adhere to some goals within its
range of physical limitation.
 This new definition posed 3 major problems: How to incorporate
and generalize observations, how to represent non-physical real
world data (say emotions), how to get knowledge about the world.
Language
Sorry no clever pictures for this
slide
 One of the most obvious applications of a
narrowly defined/capable AI system is
interpreting one language into another. “The
pen is in the box.”
 “The box is in the pen.”
 AI therefore cannot be simply a informationretrieval system, rather it must be something
more like question-answering system that not
only processes data but also amends and
draws inferences from the input.
Applied AI
 DENDRAL- chemist’s assistant in interpreting
the data from mass spectrography that
operates at the level of a Chemistry Ph.D.
 MACSYMA- mathematician's assistant that
works faster than humans in manipulating
many types of algebraic expressions.
 There is also the Sussman and Stallman
program for understanding electronic circuits.
 And many groups have dedicated their work to
producing systems to aid doctors in all types of
medical diagnosis.
Recent Attempts At AI
 Temporal Difference – TD-Gammon
 A new approach to gaming
 Rather than play out the game ahead of
time, creates a set of goals with different
weights that are adjusted each time it plays
 Capable of competing with the best players
in the world and even creating new
strategies of its own
More Recent AI
Developments
 Decision Trees
Slacking off on the pictures
 Form of classification
 Consecutive tests of input properties guide data
toward correct leaves, or create new classes when
necessary
 Learning to Reason
 Similar to Turing’s idea that intelligent systems can
mimic human development
 Program is given a period to observe an
environment then expected to carry out inductive
reasoning about that environment effectively.
Where is AI going?
 Frederick Brooks: “The quantification of information
embodies in structure.”
 Jim Gray (Microsoft): “We have been handed a puzzle:
genomes and brains work. They use much more
compact programming languages than we do.”
 Butlery Lampson: cars that don’t kill people and
automatic programming (programs that write other
programs).
 John McCarthy: computer programs with at least the
intellectual capabilities of humans
 Raj Reddy: “a computer that could read a chapter in a
book and answer the question at the end of the
chapter.”
The Feigenbaum Test
 Similar to the Turing Test.
 Perhaps a better challenge proposed by Feigenbaum
is as follows: manually program a novice view of a
given domain, write software for the system that will
read the next level in the field, augmenting new
material with that it already mastered. Allow the
system to ask questions, and answer dutifully. If
absolutely necessary, direct intervention is permissible
(i.e. hard-code the new rules) though in no more than
10 percent of the new material.
 If the computational machine is capable of keeping up
with the material such that it can pass subsequent
Feigenbaum tests then AI has been achieved.
What AI isn’t (our
opinion)
•
•
•
•
AI is not deterministic
AI isn’t an optimizing system.
AI isn’t a classification program.
AI isn’t a vast search of a
problem space.
• AI doesn’t necessarily have to
be human intelligence.
What AI is
(or might be, we really
aren’t sure)
• AI is a system that understands the
information it processes
• AI is a system that learns
• AI is a system that can interact and
manipulate its environment
• AI has initiative
• AI can always respond even when the
stimuli is outside of its experience or
internalized model of the world. (i.e. the
response can be no response, the
response cannot be the program crashing)