The combination of neural networks and fuzzy logic: the holy grail for intelligent decision making

1
Artificial Intelligence and Knowledge
ManagementAdaptive and learning behaviour of
computers

Prof dr Walter R. J. Baets Euromed
Marseille, Ecole de Management Coordinator of
EcKM The Euromed Center for Knowledge
Management Director of Notion, the Nyenrode
Institute for Knowledge Management and Virtual
Education
2
LESSONS TO LEARN FOR MANAGEMENT RESEARCH(Why are
we doing what ?) (1)
Aim We are moving into a knowledge society
Therefor we are interested in the structure of
knowledge We want to deliver decision support
based on this structure We want computers to
support human thinking
3
LESSONS TO LEARN FOR MANAGEMENT RESEARCH(Why are
we doing what ?) (2)

• Lessons to learn
• Connectionism seems humanly and organisationally
  logic
• Knowledge is built bottom-up
• Changes appear incremental (self-organised ?)
• Learning behaviour of a system (people,
  organisations, etc.)
• Dynamic, non-linear behaviour (positive
  feed-backs)

Learning and the mind-set the heart of the
decision support
4
ABOUT LEARNING, PERCEPTIONS AND KNOWLEDGE
DECISION SUPPORT Picture the mental models
(individual, shared) Follow the learning cycle
(non-linear) Allow a DSS to be learning itself
(dynamic behaviour) Other techniques to make
pictures of mind sets exist e.g.... cognitive
maps, brainstorming techniques
5
ARTIFICIAL NEURAL NETWORKS (ANN) (1)
How does the brain operate?
6
ARTIFICIAL NEURAL NETWORKS (ANN) (2)
What does an artificial neural network look like?
7
ARTIFICIAL NEURAL NETWORKS (ANN) (3)
How does an artificial neural network works (gets
trained)
W1
X1 X2 X3 X4
W2

Out-F (net)
TRESHOLD VALUE
W3
NET
W4
Output
Inputs
KNOT
8
ARTIFICIAL NEURAL NETWORKS (ANN) (4)
Comparison to other DSS techniques
(advantages) Able to simulate non-linear
behaviour Has learning behaviour Non-parametric
(no equations) Fault tolerant (can easily deal
with NAs) Seeking diversity (instead of
averages) Pattern recognition
9
EXAMPLES OF RESEARCH WITH ANNs (1)

• Market response to fast moving consumer goods (1)
• Out of sample data performance
• Continental European market with 5 (beer) brands
• Variables were market share, influenced by
• price,
• numerical distribution,
• weighted distribution,
• numerical out-of-stock,
• weighted out-of-stock,
• advertising share.
• Multiple linear regression, Multiplicatif model,
  ANNs (Backpropagation)

10
EXAMPLES OF RESEARCH WITH ANNs (2)Market
response to fast moving consumer goods (2)
MLR, MUL and ANNs performance measures for
absolute values
11
EXAMPLES OF RESEARCH WITH ANNs (3)Market
response to fast moving consumer goods (3)
12
EXAMPLES OF RESEARCH WITH ANNs (4)Market
response to fast moving consumer goods (4)
MLR, MUL and ANNs performance measures for
relative values
13
EXAMPLES OF RESEARCH WITH ANNs (5)Market
response to fast moving consumer goods (5)
14
EXAMPLES OF RESEARCH WITH ANNs (6)
Brand imaging the clients perception (1)
Well established (coffee) brand in one local
market (Holland) Output is client perception,
measured by three (fuzzy) variables Input was 35
variables A number of taste variables
(sweet/bitter) 1-7 scale A number of technical
and chemical variables numbers Kohonen maps
(non-supervised learning) Research question can
one map brand perception (in order to
manipulate in a second stage)?
15
EXAMPLES OF RESEARCH WITH ANNs (7)Brand
imaging the clients perception (2)
Results of the Kohonen network White or grey are
bad coffees black-ish are good coffees
16
EXAMPLES OF RESEARCH WITH ANNs (8)Brand
imaging the clients perception (3)
Detail a Kohonen network with Backpropagation
(more accurate) Create a model which allows one
to manipulate
17
EXAMPLES OF RESEARCH WITH ANNs (9)Brand
imaging the clients perception (4)
Results Which variables are the most crucial in
the decision process (via sensitivity analysis)?
18
EXAMPLES OF RESEARCH WITH ANNs (10)Brand
imaging the clients perception (5)Degree of
importance of the variables in the final
perceptions






Figure 9 Degree to which the different variables
influence judgement 1


19
EXAMPLES OF RESEARCH WITH ANNs (11)Brand
imaging the clients perception (6)Degree of
importance of the variables in the final
perceptions
20
EXAMPLES OF RESEARCH WITH ANNs (12)Introduction
strategy of a freight carrier in a new foreign
country (1)
Shipper survey, forwarder survey, employee survey
(total 200) 57 variables Ask for importance
of issues and performance of the carriers (and
implicitly the differences) Use self-organising
maps (Kohonen)
21
EXAMPLES OF RESEARCH WITH ANNs (13)Introduction
strategy of a freight carrier in a new foreign
country (2)
Could identify 5 virtual groups (cognitive
patterns), throughout the original 3
groups Could identify those issues which were
perceived important by all people involved Could
identify those issues where the freight carriers
are perceived to be weak
22
EXAMPLES OF RESEARCH WITH ANNs (14)Introduction
strategy of a freight carrier in a new foreign
country (3)

• Allows the company to approach the market via
• these issues, rather than via classical segments
• Gives indication that classical segmentation in
  this
• market does not seem to hold
• Gives original and additional information in
  order to design the introduction strategy

23
FUZZY LOGIC (1)
Fuzzy sets and overlapping membership-functions
24
FUZZY LOGIC (2)
Representation of the concept size using fuzzy
sets
25
FUZZY LOGIC (3)
Fuzzy rules (1)
1
26
FUZZY LOGIC (4)
Fuzzy rules (2)
27
FUZZY LOGIC (5)

• ADVANTAGES
• Smooth behaviour
• Human-like behaviour
• Natural language approach

• EXAMPLES
• Sendai Subway
• Trading systems
• Washing machines, CAM-corders, micro-waves

28
FUZZY NEURAL NETWORKS IN MANAGEMENT
Combination of the learning behaviour of neural
networks with the fuzziness and the (though
fuzzy) rules Overlapping and vague memberships
is a reality in managerial problems Fuzzy rules
is a reality in management Fuzzy and learning
behaviour is very human Pretty much to be
discovered in management sciences
29
GENETIC ALGORITHMS (1)
30
GENETIC ALGORITHMS (2)
31
GENETIC ALGORITHMS (3)
32
GENETIC ALGORITHMS (4)
33
GENETIC ALGORITHMS (5)
34
GENETIC ALGORITHMS (6)
35
GENETIC ALGORITHMS (7)
36
GENETIC ALGORITHMS (8)
37
INTELLIGENT DECISION SUPPORTQuality control
with the RWS (1)
Problem Quality management with important
motorway maintenance projects Change process
with different (and opposed) stakeholders Experim
ent with (supported) learning behaviour of
people Learning behaviour of DSSs
38
INTELLIGENT DECISION SUPPORTQuality control
with the RWS (2)
What we did and how we did it Inventory of
issues and statements Test questionnaire Field
research Training of neural networks
map (representation) Found virtual
stakeholder-groups Identified issues of
convergence and divergence
39
INTELLIGENT DECISION SUPPORTQuality control
with the RWS (3)
What are we doing now Reduce the original
picture Define the RWS aims and rules Identify
fuzzy variables out of the picture and the
aims Identify fuzzy rules Create a nice
human-machine interface Via questioning the
system allow the system to learn and give
indication on the rules Use the tool to support
learning of groups and individuals
40
POSSIBLE FUTURE DEVELOPMENTS IN AI (1)

• 1. ANNs and fuzzy logic as a tool for DSS with
• perception-data (marketing)
• Brand imaging
• Scanner data in combination with behavioural
  information
• Client perceptions (measurement of emotions)
• 2. Understanding complex systems
• Introduction strategies (fuzziness)
• Market behaviour
• Company organisational form and managerial
• structures

41
POSSIBLE FUTURE DEVELOPMENTS IN AI (2)
3. Management Information Systems controls
systems with behavioural factors 4. Support
and/or picturing learning (change) processes 5.
Machine intelligence (AI)
42
CASE BASED REASONING SYSTEM
43
RULE BASED SYSTEMS / EXPERT SYSTEMS
44
GROUP DECISION SUPPORT SYSTEMS
45
A GDSS LABORATORY
46
COGNITIVE MAPPING (1)
47
COGNITIVE MAPPING (2)
48
COGNITIVE MAPPING (3)
49
COGNITIVE MAPPING (4)
50
COGNITIVE MAPPING (5)
51
INTERNET AND INTRANET SUPPORTING THE EMERGENT
Internet supporting the emergent information
flows Intranet as a social space The role of
narratives (language) Integration of knowledge
in practice Communities-of-practice The role of
stolen knowledge
52
COMMUNITY-BASED KNOWLEDGE REFINEMENT
The case of Xerox