Transcript Artificial Intelligence

Artificial Intelligence
The Turing Test
Ian Gent
[email protected]
Artificial Intelligence
The Turing Test
Part I :
Part II:
Turing’s Imitation Game
Some sample games
from the 60’s to the 90’s
Alan M Turing, Hero
 Helped to found theoretical CS
 1936, before digital computers existed
 Helped to found practical CS
 wartime work decoding Enigma machines
 ACE Report, 1946
 Helped to found practical AI
 first (simulated) chess program
 Helped to found theoretical AI …
3
Can Machines Think?
 Computing Machinery and Intelligence




Alan M Turing
Mind, Vol LIX, Number 236 (1950)
Can be found reprinted in many places
e.g. Computers and Thought
4
Can Machines Think?
 Turing starts by defining machine & think
 Will not use everyday meaning of the words
 otherwise we could answer by Gallup poll
 Instead, use a different question
 closely related, but unambiguous
 “I believe the original question to be too meaningless
to deserve discussion”
5
The Imitation Game
 Interrogator in one room
 digital computer in another room
 person in a third room
 From typed responses only, can
interrogator distinguish between
person and computer?
 If the interrogator often guesses
wrong, say the machine is
intelligent.
 Usually done with one
machine/person at a time
6
A sample imitation game
 Turing suggests some specimen Q & A’s:
 Q: Please write me a sonnet on the subject of the Forth Bridge
 A: Count me out on this one, I never could write poetry
 Q: Add 34957 to 70764.
 (pause about 30 seconds)




A: 105621
Q: Do you play chess?
A: Yes
Q: I have K at my K1, and no other pieces. You have only K at K6 and R
at R1. It is your move. What do you play?
 (pause about 15s)
 A: R-R8 mate
7
What did Turing think?
 Turing (in 1950) believed that by 2000
 computers available with 128Mbytes storage
 programmed so well that interrogators have only a 70%
chance after 5 minutes of being right
 “By 2000 the use of words and general educated
opinion will have altered so much that one will be
able to speak of machines thinking without expecting
to be contradicted”
8
Objections and Responses
 Turing discusses responds to some objections
 Some of them can be dealt with quite quickly
 The Theological Objection
 Man has a soul, machines do not
 AT: Can we deny His power to give a soul to a machine
 Heads in the sand
 I don’t like the idea so I will ignore it
 Argument from various disabilities
 No machine can X (e.g. tell right from wrong)
 AT: Becomes a less powerful argument each day
9
Some more objections
 Lady Lovelace’s [Ada’s] objection
 computers do whatever we know how to order them to
perform , so computers cannot do anything really new
 AT: Machines constantly surprise us.
 Argument from informality of behaviour
 impossible to write down formal rules for every situation
 AT: Scientifically impossible to prove people not driven by
rules
 Argument from ESP
 Telepathy would let humans win imitation game
 AT: Put competitors in ‘telepathy-proof’ room (!)
10
Three more serious objections
 Argument from Consciousness




“No mechanism could feel pleasure, grief …
AT: Danger of Solipsism
AT: Imitation game exists now - in oral exams
Probably the most contentious objection
 Argument from continuity in the nervous system
 the brain does not operate digitally
 AT: computers can simulate continuous behavior, eg.
Statistically
11
Three more serious objections
 Mathematical Objection
 Godel’s theorem, Halting problem, etc, show that
machines cannot do ‘meta-reasoning’.
 AT: We too often give wrong answers ourselves to be
justified in being very pleased at fallibility of machines
 The mathematical, consciousness, and continuity
arguments deserve further discussion, …
 … but that’s another story
12
Some Famous Imitation Games
 1960s
ELIZA
 Rogerian psychotherapist
 1970s
SHRDLU
 Blocks world reasoner
 1980s
NICOLAI
 unrestricted discourse
 1990s
Loebner prize
 win $100,000 if you pass the test
13
The problem with ELIZA
 Eliza used simple pattern matching
 “Well, my boyfriend made me come here”
 “Your boyfriend made you come here?”
 Eliza written by Joseph Weizenbaum
 Weizenbaum so upset at credibility of users…
 his secretary wanted to use it only in private
 psychotherapists excited at prospect of Eliza-booths
 … he wrote a book to debunk the possibilities
 “Computer Power and Human Reason”
14
The problem with SHRDLU
 SHRDLU had a very limited domain
 “Look-ma-no-hands” AI
 hard to abstract lessons learnt
 natural language processing intermingled with planning, etc
 SHRDLU written by Terry Winograd
 with this and later work, he made major contributions to AI
 especially in natural language processing
15
The problem with NICOLAI
 NICOLAI was not a computer program!
 Doug Hofstadter conducted dialogue, believing NICOLAI
was electronic
 (Almost) passed the Reverse Turing Test
 Tricks like the occasional dumb answer
 but “too much cleverness in these weird responses”
16
The problem with the
Loebner Prize
 Jason Hutchens programmed the 1996 winner
 Then wrote an article
 “How to pass the Turing test by cheating” !
 “Turing’s imitation game in general is inadequate as a test of
intelligence, as it relies solely on the ability to fool people, and
this can be very easy to achieve, as Weizenbaum found.”
17
Summary: The Turing Test
 The Turing test turns a philosophical question ...
 Can Machines think?
 … Into an operational one
 Can machines play the imitation game?
 We are not near writing programs to pass the test
 The Turing test does NOT drive much AI research
 Improving the capabilities of computers DOES
18