II - Cognitive approach
A - Knowledge representation
1
1 - Representation
People understand the world by
building mental models
2
V 3.19
2 - Knowledge
Knowledge is an abstract concept. To memorize or use a piece of information, it is
necessary to represent it. We must use a language. It’s the same for a computer system, in
order to memorize and use knowledge, it is necessary to represent it as constructed as
possible.
Knowledge representation is necessary to
• keep knowledge,
• exchange knowledge (to acquire, transmit),
• modify knowledge,
• use knowledge.
No perfect representation exists, every time it’s a compromise it depends on the field dealt
with and the use.
3
V 3.19
2.1 - Different levels of knowledge
Structure
• The carburettor is a part of the engine.
Classification
• Bicycle is a kind of means of transportation.
Definition
• Prime numbers are only divisible by one and themselves.
Law
• The intensity multiplied by the voltage gives the impedance.
Rule
• If the engine stops and ignition is correct you must suspect the fuel.
Strategy
• When the patient is a child, consider infantile diseases first
4
V 3.19
2.2 - Knowledge and reasoning
The transcription of a knowledge into a symbolic language has to allow its
utilisation by a reasoning system.
Therefore, reasoning associates two overlapped aspects :
• The structure of data representing the problem
• The associated method to use the information.
The reasoning mechanism will allow to deduce dynamically new knowledge.
Here is a great difference between a database and a knowledge based system. A
database doesn’t give you anything, but the things that have been explicitly
taught.
5
V 3.19
2.3 - Knowledge representation families
6
V 3.19
First example : the factorial function
Procedural
Declarative
Factorial function of a
given number : if the
I declare that :
number is 1 then it's 1
otherwise it's the
fact(1, 1).
multiplication of this
number by the factorial of
and that :
it minus 1.
(de fact (n)
(if (= n 1)
1
(* n (fact (- n 1)))))
fact (X, Z) :fact (X-1, Y)
Z is X * Y.
Figure 2.1 - Factorial function definition, both in procedural and declarative manners
7
V 3.19
Second example :
Adjective accordance in French
Procedural
Take a noun, get its
gender G,
look for adjectives,
if there are, take
their gender Gi.
Verify if G=every Gi
Get its ...
Declarative
I declare that :
"The adjective is in
accordance with the noun".
sentence (Art,Adj,N,V) :article (Art,G,Nb),
adjective (Adj,G,Nb),
noun (N,G,Nb),
verb (V,Nb).
Figure 2.2 - Adjective accordance
8
V 3.19
3 - Representation in logics
3.1 - Propositional calculus
A proposition is an assertion (with a given syntax) that can only be valued in two ways : right
or wrong (true or false).
Example :
Man (John)
predicate
argument
Figure 2.3 - “John is a man”
represents “John is a man”, assertion that can be true or false (here, we will assume it’s
true)
A predicate can have several arguments :
• job (jean, architect)
• Temperature (machine2,lower_pump, 50)
• Following (task2, task1)
• ...
9
V 3.19
3.2 - Lack of propositional calculus
This logic isn’t powerful enough to represent properties for a group of persons.
For example how can we tell that ‘all sailors live near the sea’ without saying it for
each of them ?
10
V 3.19
3.3 - Predicate calculus
First order predicate calculus can be seen as an extension to propositional
calculus.
It has the same kernel, but adds
• The use of variables
• The ability of quantify them
Interpretation
• There exists, at least one x for which P is true :
$x P(x)
• Q is true for all the x in the domain :
"x Q(x)
Well formed formula, in predicate calculus, can have two states, just like in
propositional calculus : true or false.
11
V 3.19
Examples
$Y (dog(Y) dangerous(Y))
"XYZ (father(X,Y) father(Y,Z) grandfather(X,Z))
(It reminds me of something)
"S (job(S,sailor) live(S,sea))
12
V 3.19
Reasoning in predicate calculus
13
V 3.19
3.4 - Fuzzy logics
Uncertain reasoning, fuzzy reasoning, incomplete reasoning… all those problem
have their own logics.
Three prisoners, five hats (modal logics)
14
V 3.19
4 - Semantic networks (Quillian 68)
4.1 - Introduction
This formalism comes from work on memory organisation.
Semantic networks are especially used in natural language.
In a semantic network, knowledge is figured by a graph, nodes stand for concepts, links for
relations between concepts.
Two relations are preferred :
• Is a kind of (AKO) <
• Is a (ISA)
Bagheraa
isa
panther
ako
animal
Figure 2.4 - “Bagheraa is a panther”
15
V 3.19
4.2 - Reasoning in semantic networks
Reasoning in a semantic network is a graph extraction. It’s a structural
correspondence.
Thus, It is possible, when looking at the graph below to answer the question :
“Where does the person, john gave the book to, live ?”
John
actor
give1
object
book
title
receiver
Mary
address
Oxford
street
"Eva"
Figure 2.7 - Reasoning in a SN.
16
V 3.19
II - Cognitive approach
B - How does a
Knowledge Based System
work ?
17
6 - Rule based system
6.1- Introduction
18
V 3.19
6 - Rule based system
6.1- Introduction (2)
Each rule has the following syntax
IF <premises> THEN <conclusions>
Which means that when all premises are true, it is possible to deduce that the
conclusions are also true.
Knowledge is given to the system in a rule form (in a declarative way).
There is a separation between the knowledge (the rules) and its utilisation (the
inference engine)
19
V 3.19
6.2 - The inference engine
o
o
The inference engine
order has the same
level as the logical
representation.
Rules
Example
Engine
in proposition
calculus
IF tank-car1-empty
THEN out-of-fuel-car1
0 order
IF tank-car1 < 3
THEN out-of-fuel-car1
0+ order
IF tank(V) = 0
THEN out-of-fuel(V)
first order
Most of nowadays
inference engine are
‘first order’.
in predicate
calculus
•
Figure 2.9 - Inference engine order
20
V 3.19
6.3 - Functioning of an inference engine
21
V 3.19
6.3.1 - Forward chaining
The reasoning is led by data, this functioning is also called ‘saturation with the fact
base’.
For example in a medical diagnose we could have...
• patient-age = 15 years old
• temperature = normal
• headache = awful
…and the system would look for all possible deductions.
22
V 3.19
Milou uses forward chaining
23
V 3.19
Three steps of forward chaining
24
V 3.19
Example
•
FB : A,E, I
•
RB : 1 - if A and B then C
2 - if B then K
3 - if G and A then B
4 - if I et E then B
5 - if E then J
6 - if J and K then F
7 - if E and D then G
25
V 3.19
Functioning of an inference engine
•
Cycle
Rules
Chosen
rule
Activated
rules
Facts
base
{ } stops the engine. The final FB is : AEIBCKJF (all these facts are true)
26
V 3.19
6.3.2 - Backward chaining
It’s a type of reasoning driven by a goal (hypothesis verification)
Example
• Patient is a child
• He’s got red spots on his face
• He’s got a temperature
Verify “The child has a measles”
A goal is given, the inference engine tries to activate rules that conclude towards
this goal. The premises of these rules become new goals to be solved.
27
V 3.19
Milou uses backward chaining
28
V 3.19
Example in
backward chaining
With the same facts base as in
forward chaining, let us try to
prove the fact F.
Figure 2.13 - The AND/OR tree
used in backward chaining
29
V 3.19
(formal) Explanations
30
V 3.19
(industrial) Explanations
Peugeot-Citroën’s expert system : SIRIUS
31
V 3.19
(industrial) Explanations (2)
Peugeot-Citroën’s expert system : SIRIUS
32
V 3.19
6.3.3 - “Mixt” chaining
Some rule based systems can use
both chaining : forward and backward.
Just like a human who sometimes
tries to deduce things without aims or
who may also want to prove
something before going further in his
reasoning.
33
V 3.19
6.4 - Nota-bene
All examples given were taken from the propositional calculus. In the predicate
calculus, inferences are more difficult because a same rule can be activated more
than once.
The following rule written in the predicate calculus formalism :
• IF FATHER(x,y) AND FATHER(y,z) THEN GRANFATHER(x,z)
Will be activated …… times on this fact base :
•
•
•
•
•
FATHER(John,Louis)
FATHER(John,Paul)
FATHER(Paul,Herbert)
FATHER(Louis,Yann)
FATHER(Arthur, John)
34
V 3.19
6.5 - Advantages
Each rule is one part of the knowledge.It has to be independent from other rules.
• Modifications are easy
It’s easy to follow the reasoning built by the machine
• Explanations are given to human
The use of other logics (fuzzy for example) can introduce the idea of weight in
rules
35
V 3.19
7 - Expert Systems (E.S.)
7.1 - History
Expert Systems (ES) comes from rule based systems, but add other reasoning
techniques or knowledge representation formalisms.
The first known is DENDRAL (1965, Feigenbaum, Stanford University). It was
able to determine the chemical structure of an element from its chemical formula
and its mass spectogram.
It was an enormous program written in Fortran. Every change in knowledge was
very difficult to make, because physics knowledge and reasoning were closely
linked in the code-program.
36
V 3.19
7 - Expert system (E.S.)
7.1 - History (2)
The idea that has given birth to ES was to separate specific knowledge from the
way of using it.
This idea is the foundation of all Expert Systems.
The idea ‘came to life’ with
• MYCIN (1976, Shortliffe) blood illnesses diagnosis.
• PROSPECTOR (1979) Geology expert system
37
V 3.19
7.2 - MYCIN & E-MYCIN
38
V 3.19
7.3 - Expert System - Definition
Software able to reach performances equal to a human, on a precise area, using
a set of knowledge acquired from one or several experts of the considered area.
It is composed of (as a production system)
• an inference engine (IE)
• a fact base (FB)
• a rule base RB)
One often finds, integrated to the IE:
• a heuristic choice strategy of rules
• a meta-knowledge (ex : meta-rule for the choice of other rules).
Feigenbaum’s definition
• ES are programs conceived to reason cleverly about tasks which are believed
to require a considerable human expertise.
39
V 3.19
7.4 - Advantages
Separation between knowledge and mechanism (algorithm)
A piece of knowledge is a production rule
• knowledge can be ‘read’
• construction and evolution are easy
• explanation capacity by examining the series of rules that were used.
40
V 3.19
7.5 - Applications
ES bring assistance on problem resolution with the following characteristics:
• There is a great quantity of knowledge
• The knowledge on the area will change
• The symbolic calculation (by opposition to the numerical calculation) is the
most important.
41
V 3.19
7.6 - Teaching
Expert Systems can also be found in teaching. Not only does it know the solution
of the exercise, but it also knows the method of resolution, which enables to
explain it all.
42
V 3.19
7.7 - TD
43
V 3.19
II - Cognitive Approach
C - Knowledge acquisition,
knowledge
management
44
8 - Context
8.1 - Motivation in the choice of an ESS
The problem deals with qualitative information (not only quantitative).
Knowledge is not only knowledge of good sense, but it’s got something intuitive.
There exists one or several recognized experts, motivated and available, that
are able to solve the problem and to explain their way of reasoning.
Economic reasons are not the only ones in favour of using ESS :
• The expertise is rare or fragile (overworked expert, leaving expert,
retirement...)
• Decisions have to be made in difficult situations (quickly, when it’s hard to
decide, stress...)
45
V 3.19
8.2 - Steps of an SE development
Step 1: demonstrator
• Development of a model (demonstrator) on a part of the problem studied. The
purpose is to prove that the problem can be solved with AI techniques.
Step 2: prototype
• Development of a prototype, that integrates a complete knowledge base, but
that doesn’t care about the real environment (i.e. : without interfaces). Tests.
Step 3: integration
• Integration of the product in its final environment. Interfaces with other
software (bases on data, CAD programs, I/O processing…) that were
simulated on the prototype are now effective.
Step 4: maintenance
• In the case of knowledge that changes, updating the knowledge base, while
keeping the coherence with the operational base.
46
V 3.19
9 - Human means
An ES development needs two actors (or two groups of actors)
• an expert of the area considered
• a knowledge engineer
These two actors will work together to develop the knowledge base. The expert,
that knows perfectly his problem, will answer the knowledge engineer’s questions,
but above all, will explain his answers by giving the reasoning that drove him to
them.
47
V 3.19
9.1 - Difference between an Expert
and a Specialist
48
V 3.19
9.2 - Work
The knowledge engineer creates a link between the expert and the machine, he
mainly works with the expert, whose behaviour he observes and analysis
behaviour, and with a computer to formalise knowledge information (with objects,
rules...).
Questions,
problems
Answers,
knowledge
Formalisation
Help
Figure 2.14 - The task of a knowledge engineer
49
V 3.19
9.3 - Four steps in knowledge formalisation
(not counting methodological documents like specifications…)
Step 1 : Interviews with the expert, in order to determine the nature of reasoning.
• useful to choose the engine order, 0 for diagnosis, 1 for scheduling…)
Step 2 : Note all the descriptors and concepts used by the expert.
• for example, when using an object representation, it enables to create
classes, attributes, the hierarchy…
Step 3 : Transcribe in the chosen language (depending on the ESS), the
knowledge that is useful to solve the problem.
Step 4 : Correct and validate in presenting the same (new) problem to both the
expert and to the expert system. Compare results and explanations.
50
V 3.19
9.4 - A knowledge engineer’s
necessary qualities
Broadmindedness
Diversified and vast knowledge to approach a new area
Intellectual curiosity
Humility (keep in mind ones ignorance)
Aptitudes for communication
Skill in convincing the expert to give his knowledge
Do not try to substitute to the expert, even if his intellectual steps seem
understood
Try to show underlying and sometimes unconscious reasoning of the expert
Be able to quickly sort relevant information
Sufficient knowledge in Artificial Intelligence, in order to choose the right
representation formalism or inference engine.
51
V 3.19
9.5 - Traditional knowledge extraction
Traditional knowledge extraction made with interviews, questions and case
studies.
Several problems
• the knowledge engineer, to whom the expert explains the whole reasoning
thinks about himself as an expert,
• The expert uses an ambiguous vocabulary whose implications are numerous,
• There is a guarantee of completion for information,
• Knowledge extraction is long and expensive.
52
V 3.19
10 - Software tools
Besides the ESS, the knowledge engineer will use different software (knowledge
formalisation and browsing tools).
Nowadays, most of ESS integrate such functionalities.
53
V 3.19
10.1 - Formalisation methods
Useful to
•
•
•
•
Accelerate knowledge acquisition,
Use the same language,
Prove the completion of the knowledge base,
Increase the system maintenance
54
V 3.19
10.2 - A method... What for ?
Expert’s knowledge (in natural language)
Computer
Boolean logic
Figure 2.15 - A method ?
55
V 3.19
10.3 - Two methods
The KOD method
• Knowledge Oriented Design (Claude Vogel, CISI)
• Text transcription, linked with concepts such as taxa, acta, schema…
• Ascending methodology
The KADS method
• Knowledge Acquisition and Design Support (European Esprit project, 95
men/year)
• Descending methodology
• Coherency verification functions
• Based on ‘libraries for every ‘kind of reasoning’
• Software tools exist (from Bull, Ilog, Cap Gemini)
56
V 3.19
10.5 - At last…
Figure 2.16 - Proportion of formalised knowledge
57
V 3.19
10.6 - Conclusion
ES use get easier and easier (thanks to their great use) but we mustn’t forget that
• the knowledge base has to be built by hand, because no automatic tool exists,
• the knowledge area is very ‘small’
• the explored knowledge is often a rather superficial one, while the expert has
a depth knowledge,
• the limitation in the kinds of reasoning used by a machine (forward and
backward chaining…).
‘Second generation’ ESS now integrate
• increased structured knowledge representation,
• easier use of underlying phenomenon
• additions of other AI techniques (natural language, image processing..)
58
V 3.19
10.7 - Something to remember
59
V 3.19
© Copyright 2026 Paperzz