Chapter Seven - Computer Science

Download Report

Transcript Chapter Seven - Computer Science

CHAPTER TWELVE
The Artificial Intelligence (AI) Approach I: The Mind
As Machine
What is AI?


Intelligent Agent (IA) – complete machine
implementation of human thinking, feeling, speaking,
symbolic processing, remembering, learning, knowing,
problem solving, consciousness, planning, and decisionmaking.
AI – the computational elements of IAs
Historical Precursors



Mechanical: Calculating machines (Pascal, Leibnitz,
Newton Babbage)
Intellectual/Philosophical: Logic (Aristotle);
mathematical calculus (Leibnitz, Newton); Knowedgebased agent: (Craik); computation (Turing).
Electronic and computer: computer (Zuse, Eckart, IBM,
Intel); integrated circuit (Shockley, Kilby)
Turing’s Finite State Machine
a/b
c/d
S0
S1
g/h
e/f
S2
i/j
k/l
(A simple example)
Finite State Explanations
Sn = State (condition) definition of the system with
a number (n) indicating the specific state.
x/y =
“x” indicates what stimulus (from the external
world) is detected;
“y” what action is to be taken when “x” occurs. The
action “y” will move the state of the system to a
new state (or possibly retain the original state).
Cognitive/Behavioral Model after
Kenneth Craik
Convert to
internal
representations
External stimuli
Manipulation by
cognitive
processes.
Translate into
action
Modification of
the external
world
Computer/Cognitive Corollaries
Element
Digital computer
Turing’s Finite State
Descriptor
Central Processor
Unit (CPU)
Calculations, Logical
decisions, program
sequence control
Determines State
Transitions.
Makes cognitive
decisions (Cognitive
manipulation.)
Stores: programs,
results, temporary
results, data
Stores: state definitions
(S0,…), external
information
(“x”),Transition (IFTHEN) Rules (“x/y”)
Memory: Facts,
Cognitive Rules,
Cognitive Methods
Memory
Input/Output
Sensor information,
control of all external
system elements
(equipment)
Communication
(Bus)
Communication
between other elements
of the computer
Craik Behavioral
Model
Signals: from external
Receives sensory
sensors; to external
information (“x”), and
actuators; conversion to
provides control (“y”) to
internal representation;
external world
conversion to action
changes.
signals.
Communications with
external world
Communications with
external world
Turing and his Detractors
Category
Argument
Theological
Thinking is a function of man’s
(God-given) immortal soul.
Evaluation
This argument is a serious
restriction of the omnipotence
of the Almighty.
Mathematical
t some theorems can neither
be proved nor disproved.
no such limitations apply to
the human intellect.
Consciousness
Universal Computing Machine
can never reproduce
consciousness
This is solipsist point of view.
How do you define thinking?
Nervous system
Extrasensory percepts
The nervous system is not a
discrete-state machine. A
machine cannot mimic nervous
system behavior.
Telepathy, clairvoyance,
precognition, and psycho
kinesis cannot be replicated
by machine.
A digital computer could be
programmed to produce
results indicative of a
continuous organization
Statistical evidence for such
phenomena is, at the very
least, not convincing.
Predictive Architectures
Craik’s “predictive” has been reinterpreted by
Hawkins

Hawkins proposes an architecture based on the
neocortex. Our brains compare perceptual inputs to
expectations.

The Hawkins IA Model
ModalityIndependent
Representation
Perceptual
Objects
Perceptual
Features
Vision
Audition
Memory
Perception
Partial
Object
Representation
Emerging Technologies to Address
Capacity Challenges of “Strong AI”
Technology
Description
Potential Capacity
Nanotubes
Hexagonal network of carbon atoms
rolled up into a seamless cylinder
High density, high speed (1000
Gigahertz; thousand times a
modern computer; logical
switch size 1x10 nanometers)
Molecules
To switch states, change the energy
level of the structure within a
“rotaxane” molecule.
1011 bits per square inch
DNA
Based on human biology. Trillions of
DNA molecules within a test tube,
each performing a given operation on
differing data.
6.6 (1014) calculations per
second (cps) – 660 trillion cps
Spin (quantum computing)
Computing with the spin of electrons.
Spin is a quality of electrons within
an atom. Subject to laws of quantum
mechanics.
Mainly for memory – retains
information when power is
removed.
Light
Laser beams perform logical and
arithmetic operations.
8 trillion cps
Artificial General Intelligence (AGI)
A model envisioned by Minsky, McCarthy and
others .

A “thinking machine” with human-like “general
intelligence”.

To include: self-awareness, will, attention,
creativity as well as human qualities we take for
granted. To date, only formative thinking
characterizes AGI.

The Singularity Institute for IA
Redirects AI research and development towards
theory of AGI. Kurzweil calls its goal the “Singularity.”

Narrow AI is a context specific approach to machine
intelligence.

Goal of AGI is an intelligence that is beyond the
human level.

Approaches to AGI and its Challenges
Method
Combine narrow AI programs into an
overall framework
Challenge
Lack ability to generalize across domains.
Advanced Chatbots
The architecture of a chatbot does not support all the needs of an AGI and the
possibility of enhancing it is remote.
Emulate the brain using imaging and
other neuroscientific and psychological
tools.
We really don’t know how the brain works – software for interpretation is very
limited; the result will be a ‘human-like’ brain and the goal of AGI is to surpass
human intelligence.
Evolve an AGI; run an evolutionary
Complete models of evolution have not been fully developed; the developments
process within the computer and wait for in “artificial life” as one example of an evolutionary system have been
the AGI to evolve.
disappointing.
Use math: develop a mathematical
theory of intelligence
Current mathematical theories require unrealistic amounts of memory or
processing power.
Integrative Cognitive Architectures: a
software system with components that
We have experience from computer science and neuroscience but this is
carry out cognitive functions and connect
currently very complex and a need for extensive creative invention.
in such a way as to achieve the desired
goal.
Evolutionary Computing (EC)
Some similarity to AGI but modeled on the
principles of biological evolution.

Aims to solve real world problems: finance;
software design; robotic learning

Model and understand natural evolutionary
systems existing in: economics, immunology, ecology

A metaphor for the operation of human thought
processes – singularly germane to achieving an IA

The EC Paradigm
Select
“candidate solutions”
Evaluate fitness of solutions to
problem
Choose solutions with highest
fitness
Generate new offspring
optimum
no
yes
end
The conflict between EC/AGI and
18th Century traditions
Traditional
Conscious: we know what we think
Universal
EC/AGI
Unconscious
Partly universal
Disembodied
Embodied
Logical
Emotional
Unemotional
Emotional
Value neutral
Empathetic
Serving our own purposes and interests
Literal: fit an objective world precisely
Serving our own purposes and interests
Metaphysical
Agent-based Architectures
“every aspect of learning or other feature of
intelligence can be so precisely described that a
machine can be made to simulate it”.

IA Classifications
Acting humanly: knowledge representation,
reasoning, learning.

Thinking humanly: subsumes psychological elements
(introspection, neurological actions of brain using brain
imaging)

Thinking rationally: solve any problem described in
logical notation – exemplified by Aristotelian principles.

Acting rationally: achieve the best outcome; act best
when uncertainty exists; produce the best expected
outcomes.

Russell/Norvig Generic IAs





Simple Reflex: actions based on existing precepts (survival)
Model-based: keep track of changing precepts; maintains an
internal state that it uses to develop responses.
Goal-based: actions depend on goals; retain goal information with
desirable situations.
Utility-based: enhanced goal-based agents – add a quality factor.
Learning agents: outgrowth of Turing (universal computation); build
a learning machine and then “teach it.” (This has become a
preferred method for building state-of-the-art Ias.
Sensors and Actuators for IAs
Agent
Representative Sensor
Representative Actuators
Human
Eyes, ears, tactile, hands,
legs, mouth, nose
Hands, legs, mouth, arms
Robotic
Cognitive (software)
Cameras, infrared range
finders, tactile sensors, odor Motors and other actuators.
detectors
Keystrokes, file contents,
network packets
Display devices (optical,
audio), file outputs, packet
transmission.
Multiagent IAs
A cooperative (or noncooperative) group of IAs
capable of sophisticated information processing
activity.

Based on mechanisms that specify the kinds of
information they can exchange and their method for
doing so.

A Simple Multiagent Example: Firefighting
victim
coordinator
Medical
assistance
Fire
fighting
demolition
Fire
locator
Removal
robot
Overall Challenges to an IA
Considerable criticism of “computational” AI has come from the neuroscientific
community (Edelman and Reeke)

coding of models: programmer must find a suitable representation of the
information; what symbolic manipulations may be required; what antecedent
requirements on the representation; human cognition may not even rely on
symbolic computation at all.

categorization requirement (facts, rules): the programmer must specify a
sufficient set of rules to define all the categories that the program must support.

procedure (algorithmic processes): the programmer must specify in advance
the actions to be taken by the system for all combinations of inputs that may
occur. The number of such combinations is enormous and becomes even larger
when the relevant aspects of context are taken into account.

Crossroads
AI is emerging as a central element of cognitive
science.; methodologies lend themselves to study in :
biological modeling ; principles of intelligent behavior ;
robotics.

Numerous practical examples of IAs provide
encouraging evidence that the disciplines of
psychology, biology, computer science, and
engineering may eventually lead to a machine that
“exceeds human intelligence.”
