Bridging Artificial Intelligence, Psychometrics, and Economics: New Theory of Intelligence

1
Bridging Artificial Intelligence, Psychometrics,
and Economics New Theory of Intelligence
Rationality

• Selmer Bringsjord and Bettina Schimanski

A Science Based Approach to Decision-Making Co-s
ponsored by the Dept. of Economics and the Dept.
of Cognitive Science Friday, November 21, 2003
2
A New Kind of AI
An Answer To What is AI?

• Not an easy question to answer.
• Assume the A part is easy we know what an
  artifact is. (Webster Something created by
  humans usually for a practical purpose.)
• There is no agreement on what human intelligence
  is.
• Two notorious conferences. See The g Factor.
• But we can agree that one great success of
  psychology is testing, and prediction on the
  basis of it. (The Big Test)

3
Psychometric Artificial Intelligence

• AI is, or at least ought to be, PAI. (pronounced
  Pi, rhymes with ?)
• PAI offers a simple but radical answer
• (Naïve definition) AI is the field devoted to
  building intelligent artificial agents, i.e.,
  agents capable of solid performance on
  intelligence tests.
• Dont confuse with Some human is intelligent

4
Psychometric AI is not completely new

• Early roots PAI was implicitly entertained with
  Evans 1968 ANALOGY program
• However, PAI cannot be based on tests which
  consist solely of geometric analogies

5
Other roots

• Alan Newells 1973 You Cant Play 20 Questions
  With Nature and Win
• Does mention briefly possible PAI origins before
  his conclusion
• .. An alternative mold for such a task is to
  construct a single program that would take a
  standard intelligence test, say the WAIS or the
  Stanford-Binet.

6
Improved Definition of PAI

• Psychometric AI is the field devoted to building
  information-processing entities capable of at
  least solid performance on all established,
  validated tests of intelligence and mental
  ability, a class of tests that includes IQ tests,
  tests of reasoning, of creativity, mechanical
  ability, and so on.

7
Broad Test WAIS
Wechsler Adult Intelligent Scale
WAIS includes many sub-tests

• The Comprehension sub-test is so difficult it
  could be a motivator for the original CYC dream
• Deals with ordinary conversation

Block Design - PERI has already cracked this
Picture Arrangement - Deals with many mental
facets
PERI should specifically solve all of the WAIS -
From there, we move to all other established tests
Ex Why are the tires of automobiles made of
rubber, rather than, say, plastic?
8
PAI (continued)

• A full defense of PAI is beyond the scope of this
  short presentation, but that defense includes
  analysis of competing answers.
• One such main competing answer is due to Stuart
  Russells work on formal accounts of rationality
  from the perspectives of AI and economics.

9
Rationality and Intelligence

• Productive research in AI, both practical and
  theoretical, benefits from a notion of
  intelligence that is precise enough to allow the
  cumulative development of robust systems and
  general results. The concept of rational agency
  has long been considered a leading candidate to
  fulfill this role.

10
The agent receives percepts from the different
environments and generates a behavior or result
that in turn causes the environment to generate a
state history. The performance measure evaluates
the state history to arrive at the value of the
agent.
Figure from Stuart Russells Rationality and
Intelligence 02.
11
Rationality

• Perfect Rationality
• Calculative Rationality
• Bounded Rationality
• Asymptotic Bounded Rationality

12
Perfect Rationality
The capacity to generate maximally successful
behavior given the available information.
Fundamental inputs include E the environment
class in which the agent is to operate U the
performance measure V(f,E,U) the expected
value according to U obtained by an agent
function f in environment class E. Then a
perfectly rational agent is defined by an agent
function fopt such that
This is just a fancy way of saying that the best
agent does the best it can.
13
Calculative Rationality
The in-principle capacity to compute the
perfectly rational decision given the initially
available information.
Calculative rationality is displayed by programs
that, if executed infinitely fast, would result
in perfectly rational behavior. Unlike perfect
rationality, calculative rationality is a
requirement that can be fulfilled by many real
programs. On the other hand, calculative
rationality is not necessarily a desirable
property.
14
Bounded Optimality (BO)
The capacity to generate maximally successful
behavior given the available information and
computational resources.
Fundamental Inputs include Agent(p,M) the
agent function implemented by the program p
running on machine M PM finite set of all
programs that can be run on M Therefore the
bounded optimal program popt is defined by
Simply put, this is the notion of finding the
program that generates the optimal feasible
solution given the environment class E and
performance measure U.
15
Objection

• Bounded Optimality limits RD to the here and
  now.
• We want to be able to analyze and achieve
  successful results irrespective of the
  limitations of computers today.

Compare describing the running time of algorithms
in terms of the O( ) notation (i.e. Big O). This
provides a way to describe the complexity
independent of machine speeds and implementation
details. For this reason we investigate
Asymptotic Bounded Rationality.
16
Asymptotic Bounded Optimality (ABO)
where kM denotes a version of M speeded up by a
factor of k (or with k times more memory) and
V(f, E, U, n) is the minimum value of V(f, E, U)
for all e ? E of complexity n.
This means that the program is basically along
the right lines if it just needs a better (i.e.
faster or larger) machine to have worst-case
behavior as good as that of any other program in
all environments.
17
More Objections

• Russells approach is based on his view that
  intelligence consists of making changes to an
  environment external to the agent.
• Problem A mind able to think about things
  intelligently, independent of making changes in
  an external environment, might not only be
  intelligent but perhaps in fact ingenious?

18
Our Theory
Environment classes correspond to those for each
sub-test. The performance measure is the grading
of the results of the test. The utility is the
score. Top loop of diagram corresponds to agent
taking the test.
19
Our approach in terms of Bounded Optimality

• For PAI, one might say you cannot tackle a
  particular test with, say, a 22-state machine.
• But such a restriction might well preclude
  getting at the essence of a test or at the
  essence of what a test is a portal to.
• Our view of intelligence / rationality is
  therefore, in general, calculative.

20
Psychometric Experimental Robotic Intelligence
(PERI)

• Scorbot-ER IX
• Sony BW XC55 Video Camera
• Cognex MVS-8100M Frame Grabber
• Dragon Naturally Speaking Software
• NL (Carmel RealPro?)
• BH8-260 BarrettHand Dexterous 3-Finger Grasper
  System

21
Future work with PERI

• Thwarting Terrorism
• Picture Arrangement sub-test of WAIS
• Natural Language Processing and Generation
• Vision (Interpretation)

22
Narratological Reasoning Thwarting Terrorism
Terrorists struggle to make stories real. Their
behavior can be anticipated, and
thus thwarted. We need computers that can
imagine future events in a (twisted)
narrative. Threat anticipation (Picture
Arrangement) Predictive power done quickly .
23
(No Transcript)
24
RAIR Web and RD
Advanced Synthetic Characters
MARMML
PERI
Savant
PAI
Slate
CDs
Super Teaching
25
Questions?

• For more detailed inquiries, visit our webpage
• http//www.cs.rpi.edu/schimb/peri