Download Artificial Intelligence

Artificial Intelligence
Why Study AI?
One reason to study it is to learn more about ourselves
Introduction
Another reason is that these constructed intelligent entities
are interesting and useful in their own right
Chapter 1 & 26
AI has produced many significant and impressive products
It is clear that computers with human-level intelligence
would have huge impact on our everyday lives
What is Artificial Intelligence ?
Goals of AI
Attempts to understand and build intelligent entities
The engineering goal is to develop the concepts and
practices of building intelligent machines. Emphasis is on
There are two main ingredients in the many definition of
Artificial Intelligence:
system building
– Thought processes and reasoning (Thinking)
– Behavior and performance (Acting)
Both can either use humans as ideal or an optimal
rationality as ideal
AI Involves Many Sciences
Neuroscience (biology)
– “how does the brain work”
Cognitive science (behavior science)
– “what is intelligence”
Mechanically science (engineers)
– “how do we build robots”
Artificial Life (ecology)
– “how to build lives”
Adaptation (mathematic)
– “how can a system learn”
Rationality (economic)
– “how should we act”
The science goal is to develop concepts, theory, and
mechanisms to model intelligence. Emphasis on
understanding and automating intelligent behavior
Four Definitions of AIntelligence
Acting humanly: The Turing test approach
Thinking humanly: The cognitive modeling approach
In this course we will
talk about:
Building agents!
Thinking rationally: The laws of thought approach
Acting rationally: The rational agent approach
1
Acting Humanly:
The Turing Test
Four Faces of AI Research
Thinking rationally:
The laws of thought
approach
Acting humanly:
The Turing test approach
Acting rationally:
The rational agent approach
Acting
Rationally
Humanly
Thinking humanly:
The cognitive modeling
approach
The Interrogator
Thinking
Am I talking to a
computer or to a
human being?
Can machines think?
Can machines behave intelligently?
Problem: Turing test is not reproducible, constructive or amenable to
mathematical analysis
Thinking Humanly:
Cognitive Science
Human vs. Rational
Study human (and animal) behavior
Should computers emulate human thought?
– We don’t even know how that works!!
– Maybe we should just approximate thought by making
computers arrive at conclusions in a rational way
Should computers emulate human behavior?
– Behaviorist approach: exemplified by the Turing Test
1960s “cognitive revolution”: information-processing
psychology replaced prevailing orthodoxy of behaviorism
Requires scientific theories of internal activities of the brain
Both approaches (roughly, Cognitive Science and Cognitive
Neuroscience) are now distinct from AI
Thinking Rationally:
Law of Thoughts
Thinking vs. Acting
Does the machine need to actually think?
–
–
–
–
–
–
Playing a game of tic-tac-toe or chess
Speech recognition for airline reservations
Amazon.com’s product suggestions
Medical diagnosis system
Automated steering of a vehicle
Etc…
“Socrates is a man; all men are mortal; therefore Socrates
is mortal”
Initiated the field called “logic”
“Logicist” tradition
Problems:
– Not all intelligent behavior is mediated by logical
deliberation
– Solve problems in principal – practice. It needs guidance
2
Acting Rationally
The Foundations of AI
Rational behavior: Doing the right thing
Philosophy (428 B.C.-present)
– Logic, methods of reasoning
– Mind as physical systems
Mathematics(c.800-present)
The right thing: that which is expected to maximize goal
achievement, given the available information
– Formal representation and proof
– Algorithms, computation, (un)decidability, (in)tractability
– Probability
Economics(1776-present)
– Formal theory of rational decisions
Does not necessary involve thinking – e.g., - blinking
reflex- but should be in the service of rational action
Rational Agents
An agent is an entity that perceives and acts
This course is about designing rational agents
Abstractly, an agent is a function from percept histories to
actions:
f : P* → A
Neuroscience(1861-present)
– Memory, Cognitive processes
The Foundations of AI
Psychology(1879-present)
– Behaviorism
– Cognitive psychology
– Cognitive science
Computer engineering(1940-present)
Control Theory(1948-present)
– Simple optimal agent designs
Linguistics(1957-present)
– Knowledge representation
– Grammar
Rational Agents
For any given class of environment and tasks, we seek
the agent (or class of agents) with the best
performance
Caveat: computational limitations make perfect
rationality unachievable -> design best program for
given machine resources
History of AI
The gestation of AI
– McCulloch, Pitts (1943) artificial neurons
• artificial neurons with binary state
• can compute any Boolean function
– Hebbian learning
– SNARC: first neural network computer
– Alan Turing (1950s) Turing Test
The first AI event:
– Workshop at Dartmouth College 1956: naming “AI”
– Logic Theorist: first reasoning system
3
Early Enthusiasm (1952-1969)
“Look, Ma, no hands” era:
John McCarthy:
– LISP, Advice Taker, named the field AI.
Marvin Minsky:
– Microworlds (incl. IQ-Tests, Blocks World). p.20 fig 1.5
Arthur Samuel:
– Game of checkers
KB Systems: The Key to Power? (1969-1979)
Knowledge-based systems:
– DENDRAL: chemical analysis
• First knowledge-intensive system
Expert systems:
– MYCIN: medical diagnosis
• Rule-based system with uncertainty factors
Newell, Simon:
– GPS: General Problem Solver
– Physical symbol system hypothesis
– Imitates human problem solving
Early Enthusiasm (1952-1969)
“Look, Ma, no hands” era:
Gelernter:
– Geometry Theorem Prover
Widrow:
– Adalines
Rosenblatt:
– Perceptrons (convergence theorem)
A Dose of Reality (1966-1973)
Difficulties with early systems:
Frames:
– Taxonomic hierarchies
AI Becomes an Industry (19801988)
AI becomes an industry
– First commercially successful expert system: R1
This fuelled interest in AI, both in the academia and the industry.
Disillusionment concerning expert systems:
– Too much work to fill them with relevant rules
Overall, The AI industry boomed from a few million in 1980 to billions
of dollars in 1988. Soon after that came a period called
“AI winter” – companies suffered as they failed to deliver on the
extravagant promises – squeezed funding for research
AI Becomes an Industry (19801988)
The return of neural networks
– Reinvention of back-propagation learning
– Programs contained little or no knowledge of their
subject matter
– Intractability of many of the addressed problems
– Fundamental limitations on the basic structures
AI becomes a science
– Hypothesis + rigorous experiments = results
Intelligent agents
– Return to the “whole agent” problem
4
Recent Events (1987New systems are built on existing theories and not on brand new
ones.
Examples:
− Speech recognition (HMMs)
− Character recognition
− Planning and Reasoning (efficient representation, and
learning -> rigorous reasoning with uncertain
knowledge)
Next step:
Integrating to complete agents: SOAR, Newell (1990)
The State of the Art
Planning and scheduling: NASA
Game playing: Deep Blue (Chess playing)
– Beats human grandmasters Kasparov
Autonomous control: ALVINN
Application: Game Playing
IBM’s Deep Blue
– First AI to beat a human chess champion: Garry Kasparov,
1997
Blondie24
– Machine learning program that won a checker’s tournament
Commercial game AI
– Increase in more sophisticated AI work for “non-academic”
games
Application: Logistics Planning
Trip itineraries
– Engines such as MapQuest use AI to propose driving directions
from one location to another
Dynamic Analysis and Replanning Tool (DART)
– Used during the 1991 Persian Gulf crisis to assist in managing
military resources (over 50,000 people, vehicles, and cargo
shipments)
Airline flight scheduling
– If flights are delayed or re-routed, AI planners are used to figure
the best way to re-schedule departures and arrivals
The State of the Art
Diagnosis
Application: Speech Recognition
Airline reservation systems
– Often robust to many different voice pitches and accents
Logistics planning: DART
Robotics: HipNav
Automatic transcription
– Monitor language and content for live radio and television
– Assist in the transcribing of closed-captioned television
programs
Language understanding and problem-solving:
PROVERB
5
Application: Text Processing
Automated language translation
– Altavista’s Babelfish server
Information retrieval
– Google search engine
Conscious Computers?
Weak AI - for the hypothesis that machines could
possibly behave intelligently
– Conscious Computers? “No way!”
Strong AI - for the hypothesis that such machines would
count as having actual minds (as opposed to simulated minds)
– Conscious Computers? “Of course”
Text classification and organization
– Google news, SPAM filtering
Document summarization
– Columbia University’s Newsblaster
Alan Turing rejected the question:
– Can machines think?" - replaced it with a behavioral test
– He anticipated many objections to the possibility of
thinking machines
Application: Biology & Medicine
Diagnosis systems
– Specialists often use statistical AI tools to diagnose a patient
has a disease based on his/her symptoms
Genome analysis software
– Now that the human and other genomes are complete, AI is
used to identify new genes, infer biochemical pathways, and
compare genomes of multiple species
Philosophical Foundation
Few AI researchers pay attention to the Turing test –
concentrate on their systems' performance on practical
tasks, rather than the ability to imitate humans
Arguments for and against strong AI are inconclusive
Consciousness remains a mystery
The Ethics and Risk
Should We Develop AI?
Application: Vision
Handwriting recognition
– US Postal Service automatically sorts mail
Face recognition
Government/bank security systems
Autonomous Land Vehicle In a Neural Network (ALVINN)
– Uses camera data to automatically steer a car on a highway
at speeds up to 65 mph (from Washington, DC to San Diego
and back!)
Six potential threats to society posed by AI and related
technology:
–
–
–
–
–
People might lose their jobs
People might have too much (or too little) leisure time
People might lose their sense of being unique
People might lose some of their privacy rights
The use of AI systems might result in a loss of
accountability
– The success of AI might mean the end of human race
6
Next!
Intelligent Agents!!
7