Transcript Artificial Intelligence & Neural Networks

Chapter 13: Artificial Intelligence
Computations that make it
possible for a machine to
perceive, reason, and act in a
manner consistent with
human behavior form the
field known as artificial
intelligence.
The Turing Test
A human questioner inputs questions
and guesses which respondent is human,
based on the answers given.
If the questioner cannot tell which
respondent is human, the software
passes the test.
Chapter 13
Artificial
Intelligence
Page 139
Computer Reasoning
To simulate logical reasoning, heuristic functions are often used.
A heuristic is an artificial measure of how close the computer’s current
status is to its problem-solving goal.
F(config) =
(# of completable rows, columns, and diagonals for
X-player) – (# of completable rows, columns, and
diagonals for O-player)
if config is a non-winning
configuration

if config is an X-win
O O
-
if config is an O-win
X
X
1
-
X O O
O O
X
-
X
X O O
O X
X
2-1=1
X O O
X
X O
3-1=2
X O O
X
X
O
2-1=1
X O O
X O
X
O O
X X
X
X X
-
-
X
O O
O O
X
X
X X
X
X
3-1=2
O O O
X X
X
-
O O O
O O
X X
O X X
X
2-1=1
X
-
O O
X X
X O
3-2=1
O O
X X
X
O
2-2=0
O O
X X O
X
O O O
3-2=1
-
X
X X
O O
O X
X X
2-2=0
O O
X
X X O
2-2=0
O O
X O
X X
O O O
3-2=1
-
X
X
X
O O
O X
X
X
2-1=1
O O
X
X O X
3-1=2
O O
X X
X O
2-1=1
O O
X X
X
O
2-1=1
O O
X O
X
X
3-1=2
In the tic-tac-toe example above, when the computer is ready to make an
X-move, it uses the heuristic max{F(config), where config can be the
result of any O-move}.
Chapter 13
Artificial
Intelligence
Page 140
Expert Systems
By programming a computer with the assistance of
experts in a particular field, an expert system can be
developed to perform very specialized tasks.
Users in need of
expert assistance
complete on-line
questionnaires and
the expert system
analyzes their
responses and
assigns probabilities
to various diagnoses.
Chapter 13
Artificial
Intelligence
Page 141
Neural Networks
To simulate learning, certain multiprocessor systems, called neural
networks, have been built to “learn” to recognize particular patterns
as correct or incorrect, based upon a trial-and-error process.
In the example below, a neural network is used to teach a
computerized system how to back a truck up to a loading dock.
The physical characteristics of the
truck are programmed, with the
relationship between the steering
wheel, the tires, the cab, and the
trailer formally calculated.
Starting at some initial position, the
truck is backed up one meter at a time,
with programmed steering; the error in
the result is measured and factored into
the next attempt, until the error is zero.
Chapter 13
Artificial
Intelligence
Page 142
Natural Language Processing
Written Comprehension: How can a computer be
programmed to grasp the syntax and semantics of a
natural language?
John saw the boy in the park with
the telescope.
Question: Whose telescope is it? Answer: John’s
John saw the boy in the park with
the puppy.
Question: Whose puppy is it?
Answer: The boy’s
John saw the boy in the park with
the statue.
Question: Whose statue is it?
Answer: The park’s
Chapter 13
Artificial
Intelligence
Page 143
Speech Recognition
1. The PC sound card converts
analog sound waves spoken
into a microphone into a
digital format.
0011010101000
0101111101010
2. A software
0100110101010
acoustical
model breaks 0101011110010
the word into 1011010100110
1010101010101
phonemes.
0101010101001
0011010101000
0101111101010
0100110101010
0101011110010
1011010100110
1010101010101
0101010101001
“K”
“K”
CALM
3. A software
“AH” COMMA
“AH”
language model “M” COMPARE
“M”
compares the “P” COMPETE
“P”
phonemes to “Y” COMPLETE
“Y”
COMPUTE
words in its
“OO”
“OO”
dictionary.
“T”
“T”
4. Once the software decides on the most
likely candidate, it displays that word.
COMPUTE
Chapter 13
Artificial
Intelligence
Page 144
Robotics
Robots are programmable devices capable of manipulating objects
and performing tasks much like humans are able to do.
One of the more difficult problems when programming a robot is determining
when it is about to collide with something, when it has collided with something,
and what to do in response to a collision.
Collision Avoidance
Use scanners to determine the
robot’s proximity to other
objects, redirecting the robot
when a collision is imminent.
Collision Detection
Use sensors at
strategically located places
on the robot to determine
if a collision occurs.
Collision Reaction
Go around?
Climb over?
Bounce back?
Run away?
Drop dead?
Chapter 13
Artificial
Intelligence
Page 145
Robot Challenges
Other common human actions that are difficult to program include
propelled locomotion and manual manipulation.
Walking Gait
How can a robot be programmed
to propel itself forward on
“legs” and still maintain its
balance?
Grasping
How can a robot be programmed to
grasp part of a stack of objects,
without toppling the rest of the
stack?
Chapter 13
Artificial
Intelligence
Page 146
AI in Games
Non-player characters in games need to appear intelligent, even
though they are controlled by the game program instead of by a
game player.
Dead Reckoning
By having a game-driven
predator character react to the
anticipated position of its
player-driven prey (using the
prey’s current position and
velocity), a chase can appear
more realistic.
Flocking
By providing a group of characters with simple
goals and behaviors, a “mob mentality” can be
implemented with a minimum of code.
Chapter 13
Artificial
Intelligence
Page 147