Expert Systems ES Artificial Intelligence AI

1
Expert Systems - ESArtificial Intelligence - AI

• Mikulá Popper
• NCKU, Tainan, 2003-2004

2
PROLOGUE
Physics is facing the question What are the
qualities and characteristics of physical objects
of our universe? Biology is facing the
question What is the essence of the verity that
some physical objects are alive? Artificial
Intelligence is facing the question What is the
information-processual essence of systems capable
to rise questions as those?
3
Artificial Intelligence

• games (chess), cross-words, crypto-arithmetic
  tasks,...
• algebraic computation, geometric problems,
  calculus (e.g. differentiation, integration),
  solving mathematical tasks formulated in free
  texts, ...
• interpretation assigning meaning - of natural
  language texts, translations of such texts from
  one to another natural language domain, story
  analysis and production, ...
• active computer vision, recognition and
  understanding of scenes, situations,
• cognitive - intelligent - robots,
• diagnostics interpretation of observable (not
  wishful) features, events, behaviors, i.e.
  identification directly not visible cause based
  on (actively) perceivable phenomena,
• planning, design, construction, ...
• learning, adaptation, ...
• .....

4
AI
THEORY AND PRACTICE OF BUILDING COMPUTER PROGRAMS
THAT PERFORM INTERESTING AND USEFUL
TASKS TECHNIQUES, ALGORITHMS, AND ANALYTICAL
TOOLS THAT ARE USEFUL IN BUILDING SOPHISTICATED
EVEN ITELLIGENT COMPUTER PROGRAMS DESIGN
PROGRAMS THAT RESPOND FLEXIBLY IN SITUATIONS THAT
WERE NOT SPECIALLY ANTICIPATED BY A
PROGRAMMER CONSTRUCT PROGRAMS THAT YIELDS
PERFORMANCE THAT IF PRODUCED BY A HUMAN OR AN
ANIMAL WOULD BE CONSIDERED MANIFISTATIONS OF
INTELLIGECE (OF CERTAING KIND)
5
AI
FOR MANY TASKS CONSIDERED TO BE DIFFICULT - IT IS
EASY TO DEVELOP PROGRAMS MAKING COMPUTERS TO
SOLVE THEM AI programs schedule airlines,
control factories and nuclear power plants,
diagnose problems in complicated electronic
devices or in human body, and perform quite
usefully when translating texts from one into
other natural language, ... OTHER TASKS
CONSIDERED THAT ARE TAKEN FOR GRANTED e.g.
MUNDANE IN THEIR NATURE OR EASY TO BE SOLVED BY
EMPLOYING ONLY COMMONSENSE REQUIRE PROGRAMS
WHOSE DESIGN IS QUITE OR EVEN EXTREMELY
DIFFICULT E.g. programs capable to recognize
faces or control an automatic vacuum cleaner
without hurting pets, destroying furniture, or
distinguishing between a piece of crystal sugar
and a diamond ring.
6
Several attempts are known aiming at
replacing intuitive understanding of Artificial
Intelligence by solid logicaly and/or
mathematicaly acceptable scientific
specification/definition.

• TURING TEST ? was the first and is the most
  famous attempt by Allan Turing (1950)
• Another famous one is the Searl's
• MISTERY OF THE CHINESE ROOM ?
• No of them is considered as fully satisfying

7
DENDRAL - a practical example Feigenbaum,
Buchanan, Lederberg, 1969

• A sample is analyzed by a mas spectrograph to
  reveal its che-mical structure i.e. its
  struc-tural formula based on its sum-med
  molecular formula (e.g. C8H16NO2). The resulting
  spec-trogram provides information on the relative
  frequency of molecule sub-structures occur-rence
  having diverse mass-charge ratio.

A respective process, if naive methods are
employed, might be very much similar to the case
of crypto-arithmetic problem. It is so because
the number of all possible mechanical trials
might be in order of hundreds of thousands.
8

• Initially DENDRAL employed such a naive
  principle of infering structural formulas from
  summed mole-cular formula based on acquired mass
  spektrogram as follows
• Generate all possible chemical structures being
  consistent with the given summed molecular
  formula.
• To each generated chemical structure assign
  either a stored or just constructed corresponding
  mass spectrogram of a particular shape.
• By applying a proper method of pattern matching
  select from all so assigned spectrograms the one
  with the best match against the acquired
  spectrogram.
• The structural chemical formula related to the so
  identified spectrogram is assumed to be the one
  being searched for the analysed specimen.

9

• Such an approach is facing a
• serious problem
• HIGH COMPUTATIOAL
• COMPLEXITY!
• It is so because to each individual
• summed molecular formula,
• even in case of not large molecule,
• might be assigned an astonishing large number of
• possible spectrograms.
• Therefore an exhaustive sequence of matching
  trials is not feasible.

10
DENDRAL - a practical example Feigenbaum,
Buchanan, Lederberg, 1969

• The PROJECT team recog-nized that a naive
  mechanistic approach is not acceptable. They
  realized that specialists in organic chemistry do
  not employ pure mechanistic ap-proach, instead
  they apply available knowledge to elimi-nate
  variants out of question and to hypothesise
  rationally preferable molecular structures to be
  tested first. In this way they minimize the
  number of inevitably exhaustive (mecha-nical)
  reviewing of possibili-ties.

11
SIGNIFICANT OUTCOME OF THE DENDRAL PROJECT?
DISCOVERING POSITION OF EXPLICITKNOWLEDGE
REPRESENTATION
12

• The new sophisticated approach was based on
  employment a lot of explicitely formulated
  knowledge regarding diverse possible spectrogram
  shapes and structures corresponding to existent
  chemical materials analyzed by a mass
  spectrograph.
• Illustration
• A peak (high occurence frequency) in the
  position of mass m15 in the molecule spectrogram
  indicates presence of the metyl component (CH3)
  in it.

13
SUCH NOTIONS LEAD TO REPLACEMENT OF EXHAUSTIVE
OR STOCHASTIC BLIND SEARCH BY KNOWLEDGEABLE
GOAL-ORIENTED PROCESSES SEARCH FIRST FOR WELL
KNOWN AND FREQUENTLY OCCURING (TYPICAL) PEAK
GROUPINGS IN MASS SPECTROGRAMS, FOCUSE ATTENTION
TO SUB-STRUCTURES OF MOLECULES WITH A GIVEN
SUMMED FORMULA WHICH DO EXIST IN REALITY AND
MIGHT BE ANALYZED BY A MASS SPECTROMETRY
Illustration If it is realistic to assume that
in the analysed molecule a ketonic subgroup
(CO) might occure, then in the course of A
spectrogram analysis preference is to be given to
confirm or exclude it.
14
(CO) For that the following (production) rule
is applied IF in the mass spectrogram do exist
two positions x1 and x2 in which peaks occure
for which x1 x2 M 28 (M stands for the
molecule mass) AND in the position x1 - 28 is a
high peak, AND in the position x2 - 28 is a high
peak, AND at least in one from the positions x1,
x2 there is a high peak, THEN The molecule
contains the ketonic sub-group
15
Recognition of some particular elements and their
occurrence in the molecule structure leads to
narrowing of further possible alternatives in the
consecutive analysis and thus the problem-solving
becomes (more) feasible. This verity influenced
the previous approach in solving problems by
DENDRAL by symbolic representation of rules
corresponding to SPECIFIC KNOWLEDGE a practical
and proficient system emerged ALL RELEVANT
THEORETICAL KNOWLEDGE NEEDED FOR SOLVING
CONCERNED PROBLEMS HAVE BEEN TRANSFORMED FROM
THEIR GENERAL FORM CORREPONDING TO THE MASS
SPECTROGRAMS THE SO CALLED FIRST PRINCIPLE')
TO EFFICIENT SPECIFIC FORMS (A KIND OF
COOC/BOOK). (Feigenbaum et al. 1971)
16
THE LESSON GIVEN BY THE ILLUSTRATION IS
IMPORTANT INFORMED - KNOWLEDGEABLE SEARCH
PROCESSES HAVE IN ARTIFICIAL INTELLIGENCE PROGRAMM
ING A PARAMOUNT IMPORTANCE
17
THE VALIDITY OF THIS STATMENT REGARDS
PARTICULARLY THE SYMBOLIC DOMAIN OF ARTIFICIAL
INTELLIGENCE THE SYMBOLIC DOMAIN
REFLECT RATIONAL DELIBERATIVE MENTAL
ACTIVITIES IN THE REALM OF AWARENESS/CONSCIOUSNES
WHEN A HUMAN BEING EMPLOYES HIS KNOWLEDGE ADN
EVEN HIS/HER BELIEFS, STANDS, CONVICTIONS HE/SHE
IS CAPABLE TO EXPRESS IT EXPLICITELY
DENDRAL IS WELL ILLUSTRATING THIS CASE
18
Human beings are successful also in other kinds
of mental activities They are capable to learn
(E.G. THE MOTHER TONGUE), recognize, understand,
see, smell, walk (in streets), climb mountains,
ride cars and horses, play violine, hunt animals,
play tennis, cooperate in footbol or basketbol,
and perform countless other activities controlled
by their brain, however without capacity to
explicate how they are doing what they are doing

Most of these activities are classified
as Sub-symbolic The corresponding processes in
ai Are based on neurocomputations
19
There are nonethless another ai domains concerned
with different approaches, they employ other
computational methods appropriate to
tasks/problems having other than symbolic or
subsymbolic nature
20
SEVERAL AI APPLICATIONS COMBINE SYMBOLIC AND
SUBSYMBOLIC COMPUTATIONS NEWLY DEVISED METHODS OF
REVEALING ANALOGIES ARE CONVINCING ILLUSTRATION
OF THAT
COOPERATION (NOT ONLY IN COLLECTIVE
SPORTS) GENERALLY IN ANY POPULATION OF WHATEVER
ENTITIES (HUMAN BEINGS, ANIMALS, INSECTS ANTS,
BEES?, CEREBRAL DOMAINS, EXTREMITY JOINTS,
ETC.) LIEDS TO DISTRIBUTED ARTIFICIAL
INTELLIGENCE - DAI ALSO COOPERATING MULTIAGENT
SYSTEMS
ANOTHER AI DOMAIN IS CONCERNED WITH EVOLUTION OF
WHATEVER ENTITIES THE EMPLOYED PRINCIPLE IS
GROUNDED IN GENETIC ALGORITHMS
21
PRINCIPLES EMPLOYED IN ARTIFICIAL
INTELLIGENCE SYMBOLIC KNOWLEDGE REPRESENTATION
AND UTILISATION - SYMBOLIC COMPUTATION SUB-SYMBOLI
C KNOWLEDGE REPRESENTATION BY NEURAL NETWORKS AND
NEURO-COMPUTATION COMBINED SYMBOLIC AND
SUB-SYMBOLIC (NOT NECESSARILY NEURO)
COMPUTATION DISTRIBUTED COMPUTATION, MULTI-AGENT
COMPUTATION EVOLUTION MODELLING GENETIC
ALGORITHMS
22
DENDRAL became a source of generalized
knowledge Efficient solving tasks/problems that
by their formulation do not imply the sequence
of solving steps is based on meaningful
application of a sufficient extent general and
professionally specific knowledge. KNOWLEDGE IS
THE GROUND WHICH MAKE POSSIBLE TO EMPLOY BY
STRATEGIES CONTROLLED EXPLORATIONS OF THE PROBLEM
SPACE FOR EFFICIENT PRODUCTION OF CORRECT PROBLEM
SOLVING PROCEDURES
23
EXPERT SYSTEMS PREVAILINGLY BELONG TO THE DOMAIN
OF SYMBOLIC KNOWLEDGE REPRESENTATION AND
SYMBOLIC COMPUTATION
24
Expert systemsare informed and knowledgable

• naive noninformed blind procedures do not
  ma-nifest any intelligence, they are inefficient
  and have unacceptable computational complexity
  rarely applicable even when solving simple
  nontrivial tasks
• universal (very general) processes are
  insufficiently informed, thus weak (e.g. laws of
  gravity)
• informed processes employ prevailingly specific
  knowledge which are means for realizing specific
  goal-oriented problem-solving processes!

25
EXPERT SYSTEMS ARE KNOWLEDGE INTENSIVE SOFTWARE
PRODUCTS EMPLOYING EXPLICITELY EXPRESSED
SYMBOLICALLY REPRESENTED PROFESSIONAL
KNOWLEDGE IN COURSE OF SOLVING TASKS/PROBLEMS
26
Expert systems background

• artificial intelligence, mathematical (symbolic)
  logic, theoretical practical informatics,
  database systems, symbolic (and subsymbolic)
  representation, asociative (semantic) networks,
  psychology cognitive science, mathematics
  (including probability, graph theory, fuzzy sets)
• Programing tools KRL, FRL, Prolog, LISP, Clips,
  Java, several ES development tools (e.g.
  NexpertObject, Kappa, ART-IM, and many more)

27
Expert systems attributes - 1

• Applications solving problems with unknown
  classic algorithmic procedures (ill-formed or
  structured or incompletely specified/informed
  problems)
• Problem-solving productive non-deterministic
  procedures employing cognitive strategies that
  evenso do not guarantee finding a solution
  frequently make possible to infer required and
  proper results, though not seldom of non
  categorial and noneeqiuvocal nature being
  influenced by diverse uncertaintie
• Functional characteristics they are embodying
  opportunis-tic, (non-categorical, qualitative,
  uncertain, fuzzy, default) data driven search
  processes based on symbolic represen-tation of
  knowledge from the domain in which the
  consider-ed problem originates

28
Expert system attributes - 2

• Dynamic control relations among system
  components opportunistic data sharing, data
  exchange, and control transfer among system
  components
• Methodic of the system design, development, and
  mainte-nance autonomous development and
  modifications of individual functional system
  components, partial and gradual (step-wise)
  design, development, and modification of the
  knowledge base (the data structure comprising the
  symbols representing knowledge and strategies of
  its exploitation)
• Employment characteristics mostly on-line
  interaction with its user and/or the environment
  in which it operates, less frequently incoming
  data assessment and interpretation in the
  background when performing in interactive
  regimen with a user in many cases it is required
  that the system is capable to provide on demand
  explanation of its activities and inferred
  results, in some cases even on demand or
  automatic adjustment of its activities according
  to the needs, constraints, and user skills

29
ES (re)productive processes

• In case of well-formed tasks/problems the
  problem-solving process is based on reproduction
  of an in advance given algorithm, that is
  suggested by the problem formulation. The
  algorithm is an embodiment of a solving procedure
  a beforehand given sequence of steps. The
  implied process is considered to be of mechanical
  nature.
• In case of ill-formed tasks/problems that do not
  suggest any particular problem-solving process,
  i.e. no specific algorithm is available, the
  solving process is based on
• search algorithms when diverse operations
  (solving steps) are searched for, applied,
  tested, and when (currently) applicable, or at
  least there is no reason to exclude them, then as
  potentially proper ones they are sequenced
  (chained together),
• if the operations prove to be not applicable,
  then they are revised and replaced by other
  operation or even with their sequence.
• THIS IS THE CASE OF A PRODUCTIVE PROCESS.

30
ES productive processes

• Productive (cognitive, intelligent)
  problem-solving processes - in contrast with
  those having random or exhaustive nature employ
  knowledgeable sets of both, general and specific
  strategies, that mini-mize the computation
  complexity corresponding to problems at hand.
• This nature of productive processes makes us to
  perceive expert systems as a software system
  embo-dying knowledgeable strategies of problem
  space explorations aiming at disclosing correct,
  efficent and applicable problem-solving processes
  yielding the needed solutions.

31
ES illustrtion of knowledege application
32
GreaterThan GT LargeVessel LV SmallVessel SM
33
(No Transcript)
34
A non-monotonous process identification of the
same or possibly related lexica and syntactic
features in formulas, e.g. occurrence of the
same predicate symbols, variable symbols, or
predicates with the same arity
35
GreaterThan GT LargeVessel LV SmallVessel SV
36
GreaterThan GT LargeVessel LV SmallVessel SV
37
GreaterThan GT LargeVessel LV SmallVessel SV
38
nm (jump)
n (cycle) f(n)
NIL (end of the process)
n1 (implicit control)
39
Zbiehanie deklaratívneho programu je riadené (vo
vobecnosti) netriviálnou procedúrou. Na jej
priblíenie pouijeme túto symboliku sk stav
(stuácia) rieenia problému, g - zobrazenie stavu
na mnoinu aplikovatelných (prípustných) operácií
ok,1 t.j. situacno-akcných pravidiel2, f
funkcia výberu vhodného operátora (pravidla,
povelu, intrukcie) - ktorá vo veobecnosti môe
byt pomerne netriviálna, n poradové císlo
vybratého operátora, ktoré pri ukoncení cinnosti
je nahradené prísluným symbolom NIL. Teda
riadiaca procedúra zodpovedá zobrazeniu a
následnej aplikácii výberovej funkcie. Mono to
vyjadrit v nasledujúcej podobe
?n f(g(sk)) f(ok)
?
?NIL Výberová funkcia f, v závislosti na
dômyselnosti jej realizácie, môe na jednej
strane zodpovedat triviálnym, na druhej strane a
velmi zloitým a sofistikovaným, heuristikami
podmienovaným, procesom. Nasledujú niektoré
monosti mechanický výber najmenieho, ci
najväcieho n, náhodilý výber operácie
(akcie), výber operácie (akcie), ktorá bola
posledne preruená, teda nedokoncená, je bud zo
vetkých prípustných najveobecnejia alebo
najpecializovanejia, posledne vykonanú
najlepie doplnuje, alebo je práve jej opakom, sa
posledne vykonanej a iba ciastocne úspenej
najviac podobá, má pre uskutocnenie k dispozícii
najviac údajov alebo najlepie údaje
(spolahlivost, presnost, kategorickost,
pecifickost, senzitívnost, diskriminacná
úcinnost a pod.), výber operácie, ktorá je bud
najcastejie alebo najzriedkavejie pouívaná,
spôsobí aktiváciu bud najväcieho alebo
najmenieho poctu nadväzujúcich akcií, spôsobí
aktiváciu najlacnejích (napr. v zmysle
výpoctovej zloitosti, nárokov na doplnenie
chýbajúcich údajov a pod.) nadväzujúcich
akcií, má potenciál získat najviac nových alebo
najdiferencujúcejich informácií, sa v
analogickej situácii najcastejie
osvedcovala, vzhladom na dané kritéria, pokial
ich splnuje, zabezpecuje najrýchlejie
dosiahnutie cielového stavu, a dalie. 1
Mnoina aplikovatelných operácií implikuje vznik
nedeterminizmu ! 2 V stuáciach, pre ktoré
absentujú potrebné poznatky, táto mnoina môe
byt aj prázdna. Taký prípad vyaduje prostriedok
oetrenia vzniknutého stavu.
40
PROCEDURAL AND DECLARATIVE PROGRAMS DIFFERENCE
BETWEEN CONTROL MECHANISMS
nm (jump)
n (cycle) f(n)
NIL (end of the process)
n1 (implicit control)
DETERMINISTIC
n f(g(sk))
f(ok)
NIL
NONDETERMINISTIC
41
General schema of an EXPERT SYSTEM
42
(No Transcript)
43
(No Transcript)
44
(No Transcript)
45
(No Transcript)
46
(No Transcript)
47
(No Transcript)
48
(No Transcript)
49
(No Transcript)
50

• PRACTICING
• Design a program capable to
• figure out a triangle surface for any valid
  combination of input parameters, e.g. AB edge and
  a and ß angles
• solve simple crypto-arithmetic problems

51

• THE END TODAY