Transcript Overview - Computer Science Department

Artificial Intelligence Overview
John Paxton
Montana State University
August 14, 2003
Montana State University
A Brief Bio
• 1985 The Ohio State University, B.S.
Computer Science
• 1987 The University of Michigan, M.S.
Computer Science
• 1990 The University of Michigan, Ph.D.
Artificial Intelligence
• 2003 Montana State University –
Bozeman, Professor of Computer Science
Talk Outline
•
•
•
•
•
•
•
What is AI?
Foundations
History
Areas
Search
Knowledge Representation
Agents
What is AI?
Science Approach
1. Systems that think like humans
2. Systems that act like humans
Engineering Approach
1. Systems that think rationally
2. Systems that act rationally
Acting Humanly
• Turing Test (1950)
Thinking Humanly
• Cognitive Modelling Approach
• General Problem Solver (Newell and
Simon, 1961)
Thinking Rationally
• The laws-of-thought approach
• Syllogisms (Aristotle)
• It is difficult to code the knowledge and to
reason with it efficiently.
Acting Rationally
• Rational Agent Approach. The agent acts
to achieve the best (or near best)
expected outcome.
Foundations
• Philosophy (e.g. Where does knowledge
come from?)
• Mathematics (e.g. What are the formal
rules to draw valid conclusions?)
• Economics (e.g. How should we make
decisions to maximize payoff?)
• Neuroscience (e.g. How do brains process
information?)
Foundations
• Psychology (e.g. How do humans and
animals think and act?)
• Computer Engineering (e.g. How can we
build an efficient computer?)
• Control Theory (e.g. How can artifacts
operate under their own control?)
• Linguistics (e.g. How does language relate
to thought?)
History
• 1943-1955 Gestation.
McCulloch-Pitts, Hebb, Turing Test
• 1956. Dartmouth Conference.
• 1952-1969. Great Expectations.
Logic Theorist, GPS, Checkers, Lisp,
Microworlds (calculus)
• 1966-1973. Reality. Machine translation
(spirit == vodka), chess, intractability,
fundamental limitations (Perceptrons).
History
• 1969-1979. Knowledge-Based Systems.
Dendral (infer molecular structure)
• 1980-present. Commercial Products.
• 1986-present. Return of neural networks.
• 1987-present. Science. Hidden Markov
Models. Neural Networks. Bayesian
Networks.
• 1995-present. Intelligent Agents.
Areas
•
•
•
•
•
•
•
Agents
Artificial Life
Machine Discovery and Data Mining
Expert Systems
Fuzzy Logic
Game Playing
Genetic Algorithms
Areas
•
•
•
•
•
•
•
Knowledge Representation
Learning
Neural Networks
Natural Language Processing
Planning
Reasoning
Robotics
Areas
•
•
•
•
Search
Speech Recognition and Synthesis
Virtual Reality
Computer Vision
Search
• Missionaries and Cannibals Problem
MMM
CCC
Search
• Missionaries and Cannibals Solution
MMM
CCC
MMM
C
CC
MM
CC
M
C
M
C
MM
CC
MMM
CC
MM
CC
C
M
C
MMM
CCC
CC
MMM
C
Types of Search
• Blind Search
– Breadth-First Search
– Depth-First Search
• Informed Search
– Best-First Search
– A* Search
Breadth-First Search
MMM
CCC
MMM
CC
C
MMM
C
CC
MM
CC
M
C
Minimax Search
• Commonly used to determine which move
to make in a 2 player, strategy game.
• Deep Junior (Ban, Bushinsky, Alterman),
the reigning computer chess champion
uses minimax.
• Minimax requires an evaluation function.
Minimax Example
• Nim
4
3
2
1
1
(my move)
2
1
1
(your move)
1
(my move)
(your move)
Chess Example
maximizer
*
*
3
*
0
-5
minimizer
*
4
10
2
Knowledge Representation
• Semantic Nets
• Fuzzy Logic
• First Order Predicate Calculus
Semantic Nets
can-fly
yes
bird
is-a
is-a
is-a
robin
magpie
no
ostrich
can-fly
Fuzzy Logic
• Shaquille O’Neal is tall
1.0
tall
0.0
5’0
6’0
7’0
Fuzzy Logic
• Karim is tall (0.6) and a good teacher (0.9)
= 0.6
• Karim is tall or a good teacher = 0.9.
• Karim is not tall = 1.0 – 0.6 = 0.4
First Order Predicate Calculus
• Every Saturday is a weekend.
x Saturday(x)  weekend(x)
• Some day is a week day.
x day(x)  weekday(x)
Agents
sensors
actuators
AGENT
ENVIRONMENT
Rationality Factors
• Performance Measure
• Prior Knowledge
• Performable Actions
• Agent’s Prior Percepts
Rational Agent
• For each possible sensor sequence, a
rational agent should select an action that
is expected to maximize its performance
measure, given the evidence provided by
the sensor sequence and whatever built-in
knowledge the agent has.
Agent Terminology
• Omniscience: the outcome of its actions
are known. Impossible!
• Learning: taking actions in order to
perform better (e.g. robot vacuum cleaner)
• Autonomy: the agent relies on its own
sensors rather than built-in knowledge
Environments
• Fully observable vs. partially observable
• Deterministic vs. stochastic
• Episodic (classification) vs. sequential
(conversation)
• Static vs. dynamic
• Discrete (chess) vs. continuous (taxidriving)
• Single agent vs. multi-agent.
Types of Agents
•
•
•
•
•
•
Reflex
Model-Based
Goal-Based
Utility-Based
Learning
Combinations of the above!
Questions?