From knowledge to
intelligence – building blocks
and applications
Chitta Baral
Department of Computer Sc. & Eng.
Arizona State University
Tempe, AZ 85287
1
25th Anniversary of the 1st AAAI





This is an exciting time for AI.
Foundations are getting stronger.
Deep studies in various areas.
(Several recent graduate level texts on different subareas of AI.)
The pioneers’ early dreams are getting closer to
realization.
But we should not overlook the core and the glue
that holds together various aspects of AI.
2
What is the core and the glue of AI?

Lets start with the meaning of the word
“intelligence”.
3
Meaning of the word: “intelligence”
1. (a) The capacity to acquire and apply knowledge.
(b) The faculty of thought and reason.
(c) Superior powers of mind. See Synonyms at mind.
2. An intelligent, incorporeal being, especially an angel.
3. Information; news. See Synonyms at news.
4. (a) Secret information, especially about an actual or potential
enemy. …
Source: The American Heritage® Dictionary
4
Meaning of the word: “intelligence”
n.
1. The capacity to acquire and apply knowledge,
especially toward a purposeful goal.
2. An individual's relative standing on two quantitative
indices, namely measured intelligence, as expressed
by an intelligence quotient, and effectiveness of
adaptive behavior.
The American Heritage® Stedman's Medical Dictionary
5
Meaning of the word: “intelligence”
n.
1 a : the ability to learn or understand or to deal with new or
trying situations
b : the ability to apply knowledge to manipulate one's
environment or to think abstractly as measured by
objective criteria (as tests)
2 : mental acuteness
Merriam-Webster's Medical Dictionary
6
Meaning of the word: “intelligence”
n.
1. the ability to comprehend; to understand and profit from
experience [ant: stupidity]
2. a unit responsible for gathering and interpreting intelligence
3. secret information about an enemy (or potential enemy); “we
sent out planes to gather intelligence on their radar
coverage”
Source: WordNet ® 1.6, © 1997 Princeton University
7
The key features of an intelligent entity


it can acquire knowledge through various
means such as learning from experience,
observations, reading and processing natural
language text, etc., and
it can reason with this knowledge to make
plans, explain observations, achieve goals,
etc.
8
To learn knowledge and to reason with it



we need to know how to represent knowledge in a
computer readable format.
Knowledge Representation (KR) is important for
both reasoning and learning.
McCarthy 1959 in Programs with commonsense:
“In order for a program to be capable of learning
something it must first be capable of being told it.”
9
Importance of KR

KR is the starting point of building intelligent
entities (or AI systems), and leads to the next
steps of acquiring knowledge and reasoning
with knowledge.
10
What does KR entails?

We need languages and corresponding
methodologies to represent various kinds of
knowledge.
11
Importance of inventing suitable
knowledge representation languages

An analogy a colleague once made:
Development of a suitable knowledge
representation language and methodology
is as important to AI systems
as
Calculus is to Physics and Engineering.
12
Research agenda hinted by the analogy




What is it that makes Calculus widely used?
It is the building block results that make it useful.
Similarly, not enough to come up with a knowledge
representation language and an interpreter for it.
Need to develop the support structure, the building
block results that will facilitate the use of the
language.
13
Consolidating our thoughts!



KR plays a central role in AI.
We need to go beyond inventing a KR
language (syntax, semantics and
implementation).
We need to develop building block results
around it, so as to make it useful.
14
Historical perspective



AI pioneers (especially
McCarthy and Minsky)
realized the importance of
KR to AI.
McCarthy 1959: Programs
with commonsense
(perhaps the first paper on
logical AI).
John McCarthy
Minsky 1974: A framework
for representing
knowledge.
Marvin Minksy
15
Historical perspective – cont.

Newell and Simon:
Their early focus was
more on the
architecture of
exhibiting intelligence,
than on KR per se.
Nevertheless, their
systems did indeed
manipulate various
kinds of knowledge.
Allen Newell
Herb Simon
16
What are the properties of a good KR
language.

To start with: should be
non-monotonic

i.e., allow revision of
conclusion in presence of
new knowledge.

Hayes 1973 (Computation and
Deduction) mentions
monotonicity (calls it
“extension property”) and notes
that rules of default do not
satisfy it.
Minsky 1974 (A framework for
representing knowledge)
criticizes monotonicity of
logistic systems.

Pat Hayes
Marvin Minsky
17
Pre-1980 history of non-monotonic
logics –from Minker’s 93 survey








THNOT in PLANNER
[Hewitt in 1969]
Prolog
[Colmerauer et al. 1973]
Circumscription
[McCarthy 1977]
Default Reasoning
[Reiter 1978]
Closed World Assumption (CWA) [Reiter 1978]
Negation as failure
[Clark 1978]
Truth maintenance systems
[Doyle 1979]
1st NCAI (AAAI 1980) 1st session


Nonmonotonic logic panel
AIJ Volume 13, 1980, a special issue
18
Last twenty five years …

Many extensions, variations, and new logics,
including







Non-monotonic modal logics
Auto-epistemic logic
Conditional logics
Description logics
Probabilistic semantics for default reasoning
Probability networks (Bayes nets, structural causal
models)
AnsProlog (programming in logic with answer sets)
19
Have we invented “calculus” of KR yet?

What basic properties should it have?





have a simple and intuitive syntax and semantics;
be non-monotonic;
have the ability to represent and reason with defaults and
their exceptions;
allow us to represent and reason with incomplete
information; and
allow us to express and answer problem solving queries
such as planning queries, counterfactual queries,
explanation queries and diagnostic queries.
20
Have we invented “calculus” of KR
yet? - continued.

What properties will make it useful?




should have building block results;
should have interpreters for reasoning with the
language;
should have existing applications; and
should have systems that can learn knowledge in
this language.
21
Is AnsProlog a good candidate?
An AnsPrologor program (late 1980s) is a collection
of rules of the form:
A0 or … or Al  B1, …, Bm, not C1, …, not Cn.
where Ais, Bjs and Cks are literals.

Michael Gelfond
Jack Minker
22
Vladimir Lifschitz
Ray Reiter
Is AnsProlog a good candidate?










Its syntax uses the intuitive If-then form.
It is non-monotonic.
Can express defaults and their exceptions.
Can represent and reason with incomplete information.
Can express and answer problem solving queries.
Large body of building block results.
Various implementations: Smodels, DLV, Prolog.
Many applications built using it.
Learning systems: Progol.
Its initial paper among the top 5 AI source documents in
terms of citeseer citation.
23
How AnsProlog differs from …



Prolog: ordering matters in Prolog; can not handle cycles with
“not”; has extra-logical features; does not have disjunction and
classical negation; and is not declarative.
Logic Programming: is a class of languages and many
different semantics are proposed for “not”.
Classical Logic:




Classical logic is monotonic.
 in AnsProlog, which helps in expressing causality, is not reverse
implication.
Disjunction symbol “or” in AnsProlog is non-classical.
The negation as failure symbol “not” in AnsProlog is non-classical.
24
Transitive Closure – graph reachability

a
b
edge(a,b).



c
edge(c,d).
d
edge(d,c)
reach(a).
reach(X)  reach(Y),
edge(Y,X).
a and b are reachable
but c and d are not.
We get that conclusion
from the AnsProlog
program with edge
facts.
25
Reachability – Clark’s completion

XY edge(X,Y)  ((X=a Y =b)  (X =c  Y=d)  (X =d  Y =c))

X reach(X)  (X = a or Y (reach(Y)  edge(Y,X))

Equality axioms.

{edge(a,b), edge(c,d), edge(d,c), reach(a), reach(b) } is a model.

But so is {edge(a,b), edge(c,d), edge(d,c), reach(a), reach(b), reach(c),
reach(d) }.

Hence one can not conclude ~reach(c), ~reach(d).

Need to go beyond first order logic.
26
Canonical example of non-monotonicity –
the Drosophila of NMR

Tweety is a bird. Birds normally fly.


Does Tweety fly?
Tweety is a
Does Tweety fly?
(from http://www.vbmysql.com/albums/AroundOrlando/seaworld_penguin.sized.jpg)
27
An early formalization

fly(X)  bird(X), not ab(X).
ab(X)  penguin(X).
bird(X)  penguin(X).
bird(tweety).


We conclude fly(tweety).
But if we add



penguin(tweety).
We can no longer conclude fly(tweety)
and conclude ~fly(tweety), by virtue of CWA.
28
KR using AnsProlog – defaults and
their exceptions [Gelfond-mid90s]

Birds normally fly.

f(X) b(X), not ab(X),
not ~f(X).

Penguins are birds that
do not fly.


b(X)  p(X).
~f(X)  p(X).
About wounded birds
we do not know.

ab(X)  w(X).

29
Continuing with the birds fly example



Birds normally fly. Tweety is a bird. Penguins
normally do not fly.
Conclusion: Tweety flies.
Birds normally fly. Tweety is a bird. Penguins
normally do not fly. Tweety is a penguin.
Conclusion: Tweety does not fly.
Birds normally fly. Tweety is a bird. Penguins
normally do not fly. Tweety is wounded.
Conclusion: Tweety may or may not fly.
30
A general framework for defaults and
exceptions.

Normally a p has the
property q.

An r (which is a p) has
the property ~q.
[strong exception]

t’s are weak
exceptions.

q(X) p(X),
not ab(d, X),
not ~q(X).

~q(X) r(X).

ab(X)  t(X).
or

ab(X) not ~t(X).
31
The frame problem and situation
calculus!






Situation Calculus introduced by McCarthy and Hayes in
1969.
initially ~f, ~g. a causes f if ~g.
Is f true after executing a? Is g true after executing a?
~holds(f, s0).
~holds(g,s0).
holds(f, res(a,S))  ~holds(g,S).
How do we encode that g’s value remains unchanged?
First attempt: ~holds(F, res(A,S)  ~holds(F,S).
Causes contradiction.
Second attempt:
~holds(F, res(A,S))  ~holds(F,S), not ab(F,A,S).
ab(f,a,S)  ~holds(g,S).
32
The frame problem and situation
calculus! – cont.




What if: initially ~f, unknown g. a causes f if ~g.
Is f false after executing a?
~holds(f, s0).
holds(f, res(a,S))  ~holds(g,S).
~holds(F, res(A,S))  ~holds(F,S), not ab(F,A,S).
ab(f,a,S)  ~holds(g,S).
Above formalization will say Yes. A hasty conclusion!
Replace the ab rule by:
ab(f,a,S)  not holds(g,S).
33
Normally fluents do not change their
values – a narrative based formulation

h(F,T+1)  h(F,T), occurs(A,T), not ab(F,A,T),
not ~h(F,T+1).

~h(F,T+1)  ~h(F,T), occurs(A,T), not ab(F,A,T),
not h(F,T+1).

Translating:
a causes ~f if p1, …, pn
~h(f,T+1)  occurs(a,T), h(p1,T), …, h(pn,T).
ab(f,a,T)  occurs(a,T), not ~h(p1,T), ..., not ~h(pn,T).
34
Interpolation: maximal reasoning with
incomplete information [Baral etal. 93]

Facts about par(X,Y).

Facts about par(X,Y) and
~par(X,Y).

anc(X,Y) par(X,Y).

m_par(X,Y)  not ~par(X,Y).

m_anc(X,Y)  m_par(X,Y).

m_anc(X,Y)  m_par(X,Z),
m_anc(Z,Y).

~anc(X,Y)  not m_anc(X,Y).


anc(X,Y) par(X,Z),
anc(Z,Y).
~anc(X,Y) not anc(X,Y)
35
Planning in AnsProlog
[Subrahmanian & Zaniolo 95]
initially ~f. initially ~g.


b causes f if g.
a causes g.


goal f.

a;b is a plan that achieves the goal.
~h(f, 1).
~h(g, 1).
h(f,T+1)  oc(b,T), h(g,T).
h(g,T+1)  oc(a,T).
h(f,T+1)  h(f,T),
not ~h(f,T+1).
~h(f,T+1)  ~h(f,T),
not h(f,T+1).
 not h(f, 3).
o_oc(A,T)  oc(B,T), A  B.
oc(A,T)  act(A),
not o_oc(A,T).
36
Recap: AnsProlog is a good candidate!









Its syntax uses the intuitive If-then form.
It is non-monotonic.
Can express defaults and their exceptions.
Can represent and reason with incomplete
information
Can express problem solving queries
Large body of building block results
Various implementations: Smodels, DLV, Prolog
Many applications built using it.
Learning systems: Progol, PRISM
37
Building block results on AnsProlog
(since late 80’s early 90’s)

Subclasses



Definite, normal, extended, disjunctive.
Acyclic, signed, stratified, call-consistent, orderconsistent.
Properties





Existence of consistent answer sets.
Existence of unique answer sets.
Complexity of entailment.
Expressiveness of the entailment relation.
Relation between the literals in an answer set and the
program.
38
Building block results – continued







Transforming programs.
Equivalence of programs.
Formal relation with other languages. (classical logic,
default logic, description logics etc.)
Splitting programs – modular analysis, and computation
of answer sets.
Systematic building of programs from components:
program composition.
For which sub-class Prolog querying is sound? Complete?
See “Knowledge Representation Reasoning and
Declarative Problem Solving” for a compilation of
various building block results on AnsProlog till late 2002.
39
Relation between literals in an answer
set and the program.
Answer sets of an AnsProlog program is
analogous to models of a theory.
 If f is in an answer set A of a program P, then
there must be a rule in P such that …
 If an answer set A of a program P satisfies the
body of a rule r of P then, the head of that rule
…

40
But reasoning is complex in AnsProlog:
Complexity and Expressiveness (mid 90s)






Complexity is related to expressiveness.
Propositional AnsPrologor:
P2P-complete; captures P2P
Propositional AnsProlog:
CoNP-complete; captures Co-NP
Note: Entailment in SAT is CoNP-complete, but it does
not capture Co-NP.
Propositional AnsProlog (WFS):
P-complete; captures P on ordered databases.
Propositional AnsProlog stratified:
P-complete; captures P on ordered databases.
AnsProlog, AnsPrologor: P11-complete; captures P11
41
Reasoning Systems





Smodels (Helsinki) – 1996-97
DLV (Vienna) – 1998-99
ASSAT (Based on compiling propositional
AnsProlog to propositional logic and then
using SAT solvers) - 2003
Prolog
XSB (Stony Brook) – mid 90s
42
General Applications

Reasoning



Declarative problem solving (Answer set programming)






Reasoning with incomplete information, default reasoning.
Reasoning with preferences and priorities, inheritance hierarchies.
Planning, job-shop scheduling, tournament scheduling.
Abductive reasoning, explanation generations, diagnosis.
Combinatorial graph problems.
Combinatorial optimizations, combinatorial auctions.
Product configuration.
Involving both



Data integration
Decision support systems
Question answering
43
Some specific applications






Phylogeny construction
Abduction and preferences in linguistics
Inference of gene relation in micro-array data
Reaction control system
Question answering with respect to travel
Reasoning about cell behavior
44
REACTION
CONTROL
SYSTEM (RCS)
45
RCS/USA-Advisor (Texas Tech Univ.)





A decision support
system for shuttle
controllers.
Action: Switchon
Direct effects, plus …
Effect propagates and
affects several objects.
AnsProlog based
planner is well-suited
for this.
46
47
Signal Pathways
(from http://www.afcs.org/cm2/)
48
Reasoning about cell behavior: ASU

Biosignet-RR




Hypothetical Reasoning : side effect of drugs
Planning: therapy design
Explanation of observations: figuring out what is
wrong
Biosignet-RRH

Hypothesis generation
49
Question Answering

Text: Tweety is a bird


Question: Does Tweety fly?
Answer (average person): Yes.


Text: John took a plane from Phoenix to Pittsburgh.


Question: Where is John after that? Where is his laptop which he
always carries with him?
Answer: Pittsburgh. Pittsburgh.




Uses common-sense knowledge that birds normally fly.
Uses common-sense knowledge.
Need a repository of common-sense knowledge in a proper
form.
Early attempts: CYC (CYC’s language), OpenMind (NL)
In progress: An open source collaboration (Spring06 symp.)
50
Project Halo!





An exciting KR effort!
Halo pilot: structured around a challenge involving
71 pages of an advanced placement (AP) inorganic
chemistry syllabus. (2003-04)
Three participants in the Pilot: SRI-UT Austin;
CYCORP, Ontoprise
Ontoprise uses F-logic program.
F-logic is an object-oriented extension of AnsProlog.
51
Extensions




Sets, aggregates
Preferences
Consistency restoration
Cardinality and Weight constraints


1 { occurs(X,T) : action(X) } 1  time (T).
Probabilities
52
Putting together probability and logic

First order logics of
probability: Halpern
Joe Halpern

Probabilistic causal
models: Pearl
Judea Pearl
53
Adding probabilistic knowledge









Bayesian networks
First order logic with statistical probability (Bacchus)
Probabilistic frame based systems (Koller and Pfeffer)
Probabilistic relational models (Friedman, Getoor, Koller,
Pfeffer)
Integrating logic with probability (Halpern)
Probabilistic Horn abduction, Independent Choice logic
(Poole)
Bayesian logic programming (Kersting & De Raedt)
Probabilistic knowledge bases (Ngo & Haddaway)
Stochastic logic programming (Muggleton)
54
Adding probabilistic knowledge –cont.








PRISM (Sato)
Probabilistic logic programming (Ng &
Subrahmanian; Lukasiewicz)
Structural causal models (Pearl)
Logic programming with annotated disjunction
(Vennekens et al.)
Relational dependency networks
Relational Markov networks
Markov logic networks (Richardson & Domingoes)
P-log (Baral, Gelfond and Rushton)
55
Existing approaches …





Most do not use the full power of the logical KR
languages
The integration of logical knowledge and
probabilistic knowledge is not quite deep
All except Pearl (and P-log) do not distinguish
between conditioning/updating with observations
and actions (do).
Updates are simplistic. Do not allow updates which
is a piece of knowledge.
None allow logical reasoning in deciding the range
of a random variable.
56
Monty Hall problem: 3 doors

One of the doors below has a prize. I am
asked to pick a door.
(From http://math.ucsd.edu/~crypto/Monty/monty.html)
57
I pick door one.
(From http://math.ucsd.edu/~crypto/Monty/monty.html)
58
After I pick door 1 …


Monty opens door 3 and shows me a goat.
Should I stay with door 1 or switch to door 2.
(From http://math.ucsd.edu/~crypto/Monty/monty.html )
59
The Monty Hall problem and nontrivial conditioning

Question: Does it matter if I switch?


Which unopened door has the higher probability
of containing the prize?
The answer depends on: Whether Monty
knew where the prize is and on purpose
opened a door which has a goat.


If yes then Switch.
[increases probability from 1/3 to 2/3]
If no then probability remains 1/3.
60
Why people have difficulty answering
this?



The existing languages of probability do not
give us the syntax to explicitly express the
knowledge accompanying Monty’s action.
Lot of reasoning is done by the human being.
Halpern in his book:



gives this and other examples;
differentiates trivial conditioning versus nontrivial conditioning; and
suggests the use of protocols.
61
Representing the Monty Hall problem
in P-log. [Baral,Gelfond &Rushton 04]











doors = {1, 2, 3}.
open, selected, prize, D: doors
~can_open(D)  selected = D.
~can_open(D)  prize = D.
can_open(D)  not ~can_open(D).
random(prize).
random(selected).
random(open: {X : can_open(X)}).
pr(prize = D) = 1/3. pr(selected = D) = 1/3.
obs(selected = 1). obs(open = 2). obs(prize  2).
P(prize = 1) =? P(prize=3)=?
62
Non-trivial conditioning in P-log



We can add random atoms.
We can add rules.
Thus we can exactly specify the possibility
space of our observation.
63
Recap!




Discussed role of KR in AI.
AnsProlog: examples, applications
An important extension of AnsProlog: P-log
In terms of language design:


Adding probabilistic representation to logical KR
is a frontier.
Another frontier: High level languages
64
The Yale Shooting Problem





Published in 1986 AAAI
Received this year’s Classic Paper award.
Various encodings of an English description
did not work as expected.
The logics used in the encodings were
considered not adequate.
Lot of soul searching among researchers.
65
Impact of Yale shooting Problem

But soon other possibilities were considered

The encodings were not adequate, but correct encoding
exists.



Encoding (in logic) needs practice/experience/etc.
English description may have multiple interpretations.
How to make sure an encoding is correct?
KR moved to a more systematic approach.


Instead of the prevailing example/counter-example
approach;
Use of high level languages as a “specification”.
66
High level languages – motivation


Domain expert may not care to know the
nuances of AnsProlog or P-log.
They need an easier interface to encode
knowledge and ask query.
67
STRIPS (Fikes & Nilsson 1971):
an early HLL



Planning scenario is specified in STRIPS
syntax.
Its semantics (defined much later) can be
considered as a transition graph that is a
mapping between states due to actions.
Finding a plan is finding a path in this
diagram.
68
Post Yale Shooting – the language A
[Gelfond & Lifschitz 1992]


Similar to STRIPS, but not solely for planning.
Syntax:

Action and effect (D):

Observation (O):
Query (Q):


a causes f if p1, …, pn
executable a if q1, …, qm
initially f
f after a1, …, an
Semantics


Defines a transition diagram.
Defines when (D,O) entails Q.
69
More on A

Semantics of A

Hypothetical reasoning:


Planning:


Find a1, …, an such that (D,O) entails f after a1, …, an
Explanation


Does (D,O) entail f after a1, …, an?
Find O’ such that (D,O + O’) entails f after a1, …, an
Correctness: An encoding e in a logic L is correct if

(D,O) entails Q
iff e(D,O) entail-L e(Q).
70
Reasoning and acting in dynamic domains – role
of various HLLs in systematic design of agents

Domain (D)



Observations (O)






World -- how objects in the world are related to each other
Action Description – impact of actions on the world
what happened when, and
what is true when
(Temporal) properties of evolution (G)
Agent control/Plan/Policies (P)
Semantics: (D,O)
entails
G during/after P.
Various HLLs for D, O, G, and P.
71
Some high level languages for D and O

A [Gelfond and Lifschitz]

Action and effect:

Observation:
Query:


L [Baral, Gelfond, Provetti 1994-95]



a causes f if p1, …, pn
executable a if q1, …, qm
initially f
f after a1, …, an
Observation: f at s1
a occurs_at s.
Query:
f after a1, …, an at s.
AC [McCain & Turner; Baral 1995]

World:
p1, …, pn causes p.
72
High level language for D & O –cont.

AK [Baral & Son 1998-99]


a
determines f.
PAL [Baral, Tuan, Tran 2002]


Sensing actions:
Probabilistic information about unknown variables.
Recap: advantage of HLL for D and O:




Compact representation of the transition diagram
Have an independent (mostly automata based) semantics
“Implementation” in a logic can now be formally proven.
Counters issues raised in Yale shooting.
73
HLL for G -- evolution property of the
world



LTL -- always, eventually
CTL, CTL* -- in all paths, exists paths
P-CTL* [Baral & Zhao 04]


P-CTL* [Baral, Lifschitz & Zhao 05]


in all paths/exists path following the given policy,
In all policies/exists policies.
Last two needed to formalize properties such “trying
ones best” when actions may have nondeterministic effect.
74
HLL for control execution (P)





Action sequences
Policies (mapping from states to actions);
condition-action rules, etc.
Conditional plans
HTN [Sacerdoti 76]
Golog (alGOl in Logic), ConGOLOG etc.
[Levesque, Reiter, et. al. 96 onwards]


a1; (a2| a3 |a 4); a5; f?
HTN+Golog, HTN+GOLOGI [Baral&Tran02,04]
75
Recap on high level languages


Besides (low level) KR languages we need high
level languages/interfaces
A recap of various high level languages, most
developed in the last 15 years.



Many others exist. (e.g. Constraint Lingo.)
Most are implemented in a provenly correct way
using (lower level) KR languages.
Many applications (mentioned earlier) use HLL as
the interface language.
76
Where do we stand now?



Tremendous progress in KR – From YSP to current.
Significant body of building block results,
interpreters, query engines, and some good
applications
Great opportunities lies ahead






Project Halo/Digital Aristotle
Question Answering
Learning by Reading
Reasoning about cell behavior
Semantic Web
Exciting times!
77
Future Challenges






More of (improvements on) everything; better/faster systems
Additional formulations of reasoning terms – causes, outliers
etc.
Integration of logical KR and probability – lots of work
needed
 Perhaps P-log is a good starting point.
Similar broadening of the core in other dimensions –
preferences, weights
Integrating various kinds of inference techniques (model
enumeration, query driven, constraint propagation)
One single KR language or integration of modules in closely
related KR languages (such as AnsProlog, F-logic programs,
fragments of DL)
78
And an important and immediate
challenge (IMHO)

Need




Path




From flat KR to modular KR.
From starting from scratch to reuse of KR modules
Large knowledge bases and applications based on them
Open Knowledge Libraries
The theory and interface protocols that allows the use of modules
from the library.
Some starting points exist (CYC, KIF etc.) – need
reevaluation taking into account the recent progress
AAAI’06 Spring symposium on building of a open
knowledge library.
79
Conclusion: big picture of the talk







Knowledge representation is key to AI.
Having suitable KR languages as well as building-block
results around them is crucial for building AI systems.
AnsProlog is a good candidate to be the “Calculus” of KR
and Intelligent system building.
Need to be able to deeply integrate logical and probabilistic
knowledge.
Need high level knowledge representation languages.
Time to consider “engineering” aspects. (modules, libraries,
etc.)
Exciting time for KR (and AI in general) – the road map to
success is becoming clearer.
80
Acknowledgements for pictures
obtained from the web.















John McCarthy: www.kurzweilai.net/ bios/bio0008.html
Marvin Minsky: www.almaden.ibm.com/ cs/NPUC95/panelists.html
Pat Hayes: http://www.kuenstliche-intelligenz.de/Artikel/InterviewwithPatrickHayes.htm
Allen Newell: diva.library.cmu.edu/ Newell/
Herb Simon: http://istpub.berkeley.edu:4201/bcc/Spring2001/cio.simon.html
Vladimir Lifschitz: http://www.cs.utexas.edu/users/vl/
Michael Gelfond: http://www.cs.ttu.edu/~mgelfond/
Jack Minker: www.cs.umd.edu/ areas/db/people
Ray Reiter: www.cs.toronto.edu/ DCS/People/Faculty/reiter.html
Penguin: http://www.vbmysql.com/albums/AroundOrlando/seaworld_penguin.sized.jpg
RCS: http://www.ksl.stanford.edu/htw/dme/rcs.html
cAMP and pathway keys: http://www.signaling-gateway.org/
Monty Hall: http://math.ucsd.edu/~crypto/Monty/monty.html
Joe Halpern: www.cs.cornell.edu/ home/halpern/
Judea Pearl: http://singapore.cs.ucla.edu/LECTURE/lecture_sec1.htm
81
Thanks for the opportunity!


Thanks to my family, my students, teachers, colleagues and
sponsors!
Its been fun pursuing one (plus epsilon) of the 4 great
questions (as mentioned in Simon’s memoir of Newell)






the nature of matter,
the origins of the universe,
the nature of life, and
the workings of mind (simulating “intelligence” artificially).
and looking forward to the continuing journey.
I wish you all the best of this journey too.
82
The End!

THANKS
83