Download Frameworks for Intelligent Systems

Frameworks for
Intelligent Systems
CompSci 765
Meeting 3
Pat Langley
Department of Computer Science
University of Auckland
Outline of the Lecture
•  Computer science as an empirical discipline
•  Physical symbol systems
•  List structures and list processing
•  Reasoning and intelligence
•  Intelligence and search
•  Knowledge and intelligence
•  Implications for social cognition
2
Computer Science as an Empirical Discipline
!
In their Turing Award article, Newell and Simon (1976) make
some important claims:
• Computer science is an empirical discipline, rather than a branch
of mathematics.
• It is a science of the artificial, in that it constructs artifacts of
sufficient complexity that formal analysis is not tractable.
• We must study these computational artifacts as if they were
natural systems, forming hypotheses and collecting evidence.
They propose two hypotheses based on their founding work in
list processing and artificial intelligence.!
3
Laws of Qualitative Structure
!
The authors introduce the idea of laws of qualitative structure,
which are crucial for any scientific field’s development:
• The cell doctrine in biology
• Plate tectonics in geology
• The germ theory of disease
• The atomic theory of matter
They propose two such laws, one related to mental structures
and another and the other to mental processes.
4
Physical Symbol Systems
!
Newell and Simon’s first claim, the physical symbol system
hypothesis, states that:
• A physical symbol system has the necessary and sufficient
means for general intelligent action.
They emphasize general cognitive abilities, such as humans
exhibit, rather than specialized ones.
This is a theoretical claim that is subject to empirical tests, but
the evidence to date generally supports it.!
5
More on Physical Symbol Systems
!
What do Newell and Simon mean by a physical symbol system?
• Symbols are physical patterns that are stable unless modified.
• Symbol structures or expressions are organized sets of symbols.
• A physical symbol system create, modifies, copies, and destroys
symbol structures in order to:
• Maintain structures that designate other objects or processes;
• Interpret expressions that designate such processes.
These ideas are agnostic about the nature of physical patterns;
they can reside in neurons, in silicon chips, or on paper.!
6
Development of the Hypothesis
!
Four historical developments during the 20th Century led to the
physical symbol system hypothesis:
• Studies in formal logic and symbol manipulation
• Turing machines and digital computers
• The concept of a stored program
• List processing and languages like IPL and Lisp
Later work on computer systems that designated and interpreted
symbol structures built on these advances.!
7
What is List Processing?
!
Newell, Shaw, and Simon introduced list processing in 1956;
this paradigm involved three key ideas:
• Dynamic memory structures and mechanisms to alter them
• Data types and operations for different types
• Designation and manipulation of symbol structures
These support abstraction of structures and processes beyond the
specific hardware on which they are implemented.
They demonstrated these ideas in IPL, the first list-processing
language, although Lisp soon became more widely used.!
8
Why is List Processing Important?
!
This new framework was very important to AI’s development
because it could:
• Encode arbitrarily complex structural descriptions;
• Create new structural descriptions dynamically;
• Use such structures to designate other structures; and
• Interpret these structures to produce behavior.
As we will see, each of these abilities plays a crucial role in
the construction of intelligent systems.!
9
Lists and List Structures
!
List-processing techniques can encode very complex structural
descriptions using:
• Symbols such as on, A, and B ;
• Lists such as (on A B) and (eats agent John object soup) ;
• List structures such as (goal me (not (on ?any B))) .
We can use such structures to represent an agent’s beliefs and
goals, as well as its rules and other forms of knowledge.
Most classical AI systems have depended on list structures to
encode such symbolic content.
10
Three Metaphors for Intelligence
Most research on AI adopts one of three distinct metaphors for
understanding the mind:
•  Intelligence as multi-step reasoning
•  Intelligence as search through a maze
•  Intelligence as retrieval of knowledge from memory
None of these metaphors is right or wrong, but each one offers
important insights into the nature of mental processing.
11
Metaphor 1: Reasoning and Intelligence
One metaphor for intelligence emphasizes the human ability to
carry out complex reasoning, including:
•  Deduction, which draws logical conclusions from givens;
•  Abduction, which finds plausible accounts of observations;
•  Analogy, which maps new situations onto known ones.
AI researchers have developed systems that exhibit each of these
cognitive capabilities.
Despite differences in operation, they all operate over structured,
relational encodings, stated as list structures.
12
Automated Reasoning
Some of the earliest AI research dealt with automated reasoning:
• Newell, Shaw, Simon’s (1957) Logic Theorist: The first AI system,
based on studies of human reasoning, proved theorems in logic;
• Gelernter’s (1959) Geometry Theorem Machine: Proved geometry
theorems with diagrams, some requiring constructions;
• Slagle’s (1963) SAINT: Solved final exam problems in symbolic
integration from a first-year Calculus course.
Automated reasoning methods are now mature enough to appear
in software like Maple and to be used in hardware verification.
Propagation of values in Excel is an even more widespread use. 13
SHRDLU (1970)
SHRDLU was an early AI system that
interacted with users in natural language.
It inferred sentence meanings to answer
queries and executed complex commands.
Although limited in scope, SHRDLU had
many features of a cognitive system.
Person: Pick up a big red block.
Computer: OK.
Person: Grasp the pyramid.
Computer: I don't understand which
pyramid you mean.
Person (changing their mind): Find a block
which is taller than the one you are
holding and put it into the box.
Computer: By "it", I assume you mean the
block which is taller than the one I am
holding.
Computer: OK.
Person: What does the box contain?
Computer: The blue pyramid and the blue
block.
Person: What is the pyramid supported by?
Computer: The box.
Person: How many blocks are not in the
box?
Computer: Four of them.
Person: Is at least one of them narrower
than the one which I told you to pick up?
Computer: Yes, the red cube.
14
Metaphor 2: Intelligence and Search
A second metaphor for intelligence rests on an analogy with
finding one’s way through a physical maze.
This view is commonly adopted
for problem solving tasks.
These require one to find some
sequences of steps that solves
an unfamiliar problem.
Instances include generating new
plans, schedules, and designs.
The search view is not antithetical to the reasoning metaphor,
but they emphasize different aspects of intelligence.
15
Problem Solving and Symbol Systems
!
Human intelligence includes the ability to solve novel problems.
But how can we find solutions to problems when we do not
already know the answers?
• This is the Meno Paradox that Plato first posed.
Fortunately, we can separate generators for candidate solutions
from mechanisms for testing them.
• This division eliminates the apparent paradox.
But it requires the ability to represent candidate solutions and
to search through the resulting problem space.!
16
The Heuristic Search Hypothesis
!
This insight led Newell and Simon to propose their heuristic
search hypothesis:
• A problem solver represents candidate solutions in terms of
symbol structures;
• Problem solving involves a search process that generates
and modifies these structures;
• The problem solver tests candidates to determine whether
they are acceptable.
This process relies on heuristics because, in practice, one cannot
search most problem spaces exhaustively.
17
Search and the Tower of Hanoi
Consider a puzzle that is known as the Tower of Hanoi, which
involves three pegs and N disks.
The task is to move all disks from the initial peg to another peg,
which involves problem solving and search.
18
Problem Space for Tower of Hanoi
Starting from a single
state, we can apply
operators repeatedly
to generate the entire
problem space, or
we can explore it
more selectively.
Initial state
Goal state
19
The Logic Theorist
!
In 1956, Newell, Shaw, and Simon developed the Logic Theory
Machine, the first running AI program.
• This system proved theorems in propositional logic, finding
different proofs from those of Russell and Whitehead.
• The Logic Theorist demonstrated the use of list processing,
heuristic search, and goal-directed reasoning.
• This was first of an entirely new breed of computer programs.
It would difficult to overestimate the system’s impact on the
new field of artificial intelligence.!
20
AI Planning Systems
One important use of search mechanisms is to generate a plan
for achieving some goal.
AI planning methods have become mature enough to support a
variety of applied systems, including:
•  Automobile navigation systems that generate routes to follow;
•  Orbitz and other travel sites, which propose airline itineraries;
•  DART, which generated logistical plans for the US military;
•  The Hubble space telescope, Mars rovers, and even copiers.
These systems are much less flexible than human planners, but
they find solutions that people might overlook.
21
Game-Playing Programs
Early work on games like chess, one of AI’s original challenge
problems, led to many insights about representation and search.
•  In 1957, Herbert Simon predicted that “within ten years, a computer
will be the world’s chess champion”.
•  Simon’s prediction was too optimistic, but only his timing was off.
•  In 1997, Deep Blue, a hardware-aided chess program that searched
deeper than humans, beat Gary Kasparov, the human champion.
Similar results have occurred for checkers, backgammon, and
other complex games that humans find challenging.
However, these systems are tuned for specific games and lack
the generalized abilities of human players.
22
Metaphor 3: Knowledge and Intelligence
A third metaphor for intelligence emphasizes people’s ability to
draw upon large amounts of knowledge in memory about:
•  The structure and behavior of physical environments;
•  Topics from science, engineering, arts, and humanities; and
•  Relations and events that arise in social interactions.
AI researchers have developed systems that store, retrieve, and
use knowledge in each of these areas.
This paradigm led to the first widespread applications of AI
technology in commerce and industry.
23
Expert Systems
The expert systems movement built upon insights about the role
of knowledge in human expertise.
This paradigm encodes knowledge as if-then rules or as stored
cases that are matched to draw conclusions.
•  Hundreds of such expert systems have been deployed since the
1980s, making or saving substantial money.
•  However, they are costly to maintain as conditions change, and
their reasoning is often routine and shallow.
Successes of machine learning and data mining are best viewed
as the automated construction of expert systems.
24
The TurboTax System
One widely adopted expert system is TurboTax, which helps its
users prepare their US tax returns.
TurboTax has been purchased by millions of customers, reducing
their time and effort to file taxes.
25
Case-Based Systems
An alternative approach to mimicking expert behavior involves
case-based systems that:
•  Store / index specific instances or cases in long-term memory.
•  Upon encountering some new problem, retrieve a similar case.
•  Use the retrieved case to draw inferences or propose a response.
•  Adapt this case to the new situation if needed (optional).
Case-based systems are arguably easier to construct, understand,
and maintain than rule-based expert systems.
They have also been quite successful in commercial settings,
including interactive ones like help desks.
26
Recommender Systems
In the 1990s, increasing availability of Web content, including
products on the Web, led to recommender systems that:
•  Propose, rank, or otherwise present selected items that they
predict will interest the user.
•  Utilize profiles for users that describe their preferences, which
are often learned from explicit or implicit feedback.
•  Amazon, Tivo, and many other companies use techniques of
this sort to increase sales and customer satisfaction.
Recommender systems have shallow user models and limited
interaction abilities, but they are used widely.
27
Summary Remarks
!
We have reviewed some broad claims that appear relevant to
interactive cognitive systems:
1.  Computer science and AI are empirical disciplines that study
the behavior of computational artifacts.
2.  Physical symbol systems are necessary and sufficient for
intelligent behavior (physical symbol system hypothesis).
3.  Problem solving involves heuristic search in a space of states
encoded as symbol structures (heuristic search hypothesis).
4.  AI has explored two other main metaphors for intelligence:
as reasoning and as expertise.
These insights and frameworks have become so basic that many
AI researchers do not realize they were once unknown.
28
Implications for Social Cognition
!
How are these ideas relevant to the social cognition hypothesis
proposed earlier? They suggest that we should:
• Develop computational artifacts to understand the social
character of intelligence;
• Represent our agents’ beliefs, goals, and knowledge using
symbolic structures;
• View social understanding and planning in terms of heuristic
search, reasoning, and knowledge use.
However, interaction raises new issues; e.g., in social planning,
should problem states include models of others’ beliefs / goals?
29
Reading Assignments
We will have two reading assignments for the next meeting:
•  Young, R. M. (2001). Production systems in cognitive
psychology. In N. J. Smelser & P. B. Baltes (Eds.), International
Encyclopedia of the Social and Behavioral Sciences. Pergamon.
•  Jones, G., & Ritter, F. E. (2003). Production systems and rulebased inference. Encyclopedia of Cognitive Science. London:
Macmillan.
These papers provide brief overviews of production systems,
an important class of formalisms for high-level cognition.
Please read both papers and come prepared to discuss them.
30
End of Lecture!
31