Dr Kevin Voges

1
__________________________________ Applications
of Artificial Intelligence Techniques in Business
__________________________________
Dr Kevin Voges Department of Management University
of Canterbury Christchurch New Zealand January
2006
2
What is Artificial Intelligence? Brief
history Applications Data analysis using rough
sets Agent based market simulation Conclusions

3
What is Artificial Intelligence? the branch
of computer science that is concerned with the
automation of intelligent behavior (Luger
Stubblefield, 1998) - advanced computer
science - based on sound theoretical principles
- automation - produces a collection of
problems and methodologies for their
solution - produces working software that
produces useable results
4
What is Artificial Intelligence? the branch
of computer science that is concerned with the
automation of intelligent behavior (Luger
Stubblefield, 1998) - intelligent behaviour -
representing knowledge - planning ahead -
recognizing patterns - understanding natural
language - negotiating - intelligence as an
adaptive process (Piaget, 1978) - biological and
social models of intelligence - embodied AI
5
Brief History Charles Babbage - 1837 - built a
mechanical difference engine - designed
mechanical analytical engine - similar
underlying assumptions as modern computers
George Boole - 1854 - laws of thought (Boolean
Algebra) - demonstrated that a very simple
formal system could capture the full power of
logic
6
Modern Era Alan Turing - worked on Enigma
Machine during WW2 - developed theoretical
principle for universal computing machine 1948
- Unpublished Manuscript - investigated the
question as to whether it is possible for
machinery to show intelligent behaviour - some
concepts, including networks of artificial
neurones, only became widely available after
reinvention by other researchers. 1956
Dartmouth Conference on Artificial Intelligence
7
Modern Artificial Intelligence has many
sub-disciplines - Reasoning with uncertain or
incomplete information - Bayesian reasoning,
fuzzy sets, rough sets, etc. - Knowledge
representation - frames, scripts - conceptual
graphs (networks of beliefs) - association
rules - Connectionist approaches - neural
networks - self-organizing maps - Adaptive and
emergent approaches - genetic algorithms,
evolutionary algorithms - simulated annealing -
artificial immune systems - Other techniques -
support vector machines - agent-based
approaches - Hybrid techniques
8
Neural Networks Most widely reported technique
in marketing literature Aincough and Aronson
(1999) - scanner data analysis Alon, Qi and
Sadowski (2001) - forecasting aggregate retail
sales Curry and Moutinho (1993) - modelling
consumer response to advertising stimuli Law and
Au (1999) - forecasting Japanese demand for
travel to Hong Kong
9
__________________________________
Applications __________________________________
Marketing Data Mining E-commerce Production
and Operations Finance Decision-making General
Management
10
__________________________________ Specific
Applications __________________________________
Data Analysis using Rough Sets Agent Based
Market Simulation
11
__________________________________ ROUGH
SETS __________________________________
Very brief history Basic concepts
Information tables and decision tables
Partitions Mappings Inconsistent information
tables Rough (approximate) sets Application
Rough clustering
12
__________________________________ Very Brief
History __________________________________
Zdzislaw Pawlak Warsaw University of
Technology (1982) Rough sets. (1984) Rough
classification. (1991) Rough sets Theoretical
aspects of reasoning about data. Good
introduction in Munakata (1998)

13
__________________________________ Information
Tables and Decision Tables Information
Table __________________________________
14
__________________________________ Information
Tables and Decision Tables Information
Table __________________________________
Object Attributes (Record / (Variables) Case)
15
__________________________________ Information
Tables and Decision Tables Information
Table __________________________________
Object Attributes (Record / (Variables)
Case) Gender Income
16
__________________________________ Information
Tables and Decision Tables Information
Table __________________________________
Object Attributes (Record / (Variables)
Case) Gender Income a male low b male
low c male medium d female medium e female
high f female high
17
__________________________________ Information
Tables and Decision Tables Decision
Table __________________________________
Object Attributes Condition Decision Gender
Income Purchase a male low no b male
low no c male medium yes d female
medium no e female high yes f female
high yes
18
__________________________________ Information
Table __________________________________
G In P a m l n b m l n c m m y d f m n e f h y
f f h y
19
__________________________________
Partitions __________________________________
A partition is the set of disjoint subsets,
U Subsets are called -
Equivalence Classes - Blocks
G In P a m l n b m l n c m m y d f m n e f h y
f f h y
20
__________________________________
Partitions __________________________________
A partition is the set of disjoint subsets,
Gm, Gf Gm a, b, c Gf d, e, f
U Subsets are called - Equivalence
Classes - Blocks
G In P a m l n b m l n c m m y d f m n e f h y
f f h y
21
__________________________________
Partitions __________________________________
A partition is the set of disjoint subsets,
Gm, Gf Gm a, b, c Gf d, e, f
Where Gm U Gf U and Gm Ç Gf
Subsets are called - Equivalence Classes -
Blocks
G In P a m l n b m l n c m m y d f m n e f h y
f f h y
22
__________________________________
Partitions __________________________________
Partition induced by Income, Inl, Inm , Inh
Where Inl U Inm U Inh U and Inl Ç
Inm Inm Ç Inh Inl Ç Inh
G In P a m l n b m l n c m m y d f m n e f h y
f f h y

23
__________________________________
Partitions __________________________________
Partition induced by Income, Inl, Inm , Inh
Where Inl U Inm U Inh U and Inl Ç
Inm Inm Ç Inh Inl Ç Inh
G In P a m l n b m l n c m m y d f m n e f h y
f f h y
On the basis of Income, objects a and b are
indiscernible, objects c and d are
indiscernible, and objects e and f are
indiscernible

24
__________________________________ Partitions of
Two Attributes __________________________________
Partitions induced by Gender and Income, X1 (
G-m, In-l ) a, b X2 ( G-m, In-m ) c
X3 ( G-f, In-m ) d X4 ( G-f, In-h
) e, f
G In P a m l n b m l n c m m y d f m n e f h y
f f h y

25
__________________________________ Partitions of
Two Attributes __________________________________
Partitions induced by Gender and Income, X1 (
G-m, In-l ) a, b X2 ( G-m, In-m ) c
X3 ( G-f, In-m ) d X4 ( G-f, In-h
) e, f Where X1 U X2 U X3 U X4 U
G In P a m l n b m l n c m m y d f m n e f h y
f f h y

26
__________________________________ Partitions of
Two Attributes __________________________________
Partitions induced by Gender and Income, X1 (
G-m, In-l ) a, b X2 ( G-m, In-m ) c
X3 ( G-f, In-m ) d X4 ( G-f, In-h
) e, f Where X1 U X2 U X3 U X4 U and
X1 Ç X1 X2 Ç X3 X1 Ç X2 X2 Ç
X4 X1 Ç X3 X3 Ç X4
G In P a m l n b m l n c m m y d f m n e f h y
f f h y

27
__________________________________ Partition of
Decision Attribute _______________________________
___
G In P a m l n b m l n d f m n c m
m y e f h y f f h y Note change of order
Partition induced by Purchase Decision, Pn, Py
Where Pn U Py U and Pn Ç Py
.

28
__________________________________ Partition of
Decision Attribute _______________________________
___
G In P a m l n b m l n d f m n c m
m y e f h y f f h y
Partition induced by Purchase Decision, Pn, Py
Where Pn U Py U and Pn Ç Py
Subsets are called concepts.

29
__________________________________
Mappings __________________________________
Rough sets theory is interested in finding
mappings from the partitions induced by the
condition attributes to the partitions induced by
the decision attribute(s). These mappings can be
presented as rules.
G In P a m l n b m l n d f m n c m
m y e f h y f f h y

30
__________________________________
Mappings __________________________________
Rough sets theory is interested in finding
mappings from the partitions induced by the
condition attributes to the partitions induced by
the decision attribute(s). These mappings can be
presented as rules. eg. X - Same gender and
same income - results in four equivalence
classes If X1 a, b then Pn a, b, d
If X2 d then Pn a, b, d If X3
c then Py c, e, f If X4 e, f
then Py c, e, f
G In P a m l n b m l n d f m n c m
m y e f h y f f h y

31
__________________________________
Mappings __________________________________
Rough sets theory is interested in finding
mappings from the partitions induced by the
condition attributes to the partitions induced by
the decision attribute(s). These mappings can be
presented as rules. eg. - If gender is male and
income is low, then no purchase.
G In P a m l n b m l n d f m n c m
m y e f h y f f h y
32
__________________________________
Mappings __________________________________
Rough sets theory is interested in finding
mappings from the partitions induced by the
condition attributes to the partitions induced by
the decision attribute(s). These mappings can be
presented as rules. eg. - If gender is female
and income is medium, then no purchase.
G In P a m l n b m l n d f m n c m
m y e f h y f f h y
33
__________________________________
Mappings __________________________________
Rough sets theory is interested in finding
mappings from the partitions induced by the
condition attributes to the partitions induced by
the decision attribute(s). These mappings can be
presented as rules. eg. - If gender is male and
income is medium, then purchase.
G In P a m l n b m l n d f m n c m
m y e f h y f f h y
34
__________________________________
Mappings __________________________________
Rough sets theory is interested in finding
mappings from the partitions induced by the
condition attributes to the partitions induced by
the decision attribute(s). These mappings can be
presented as rules. eg. - If gender is female
and income is high, then purchase.
G In P a m l n b m l n d f m n c m
m y e f h y f f h y

35
__________________________________ Inconsistent
Information Tables _______________________________
___
G In P a m l n b m l n d f m n c m
m y e f h y f f h y g f h n
Previous rule - If gender is female and income is
high, then purchase. This rule is no longer
true. An inconsistent information table contains
objects whose condition attributes are the same,
but lead to different concepts. To deal with
inconsistent information tables, we use rough
sets theory.

36
__________________________________ Rough
(Approximate) Sets _______________________________
___
A rough set is formed from two sets Lower
approximation - contains objects that are
definitely in the set. The remaining objects
from the complete (universal) set are either
definitely not in the set, or their set
membership is unknown. The set of objects whose
membership is unknown is called the boundary
region. Upper approximation - the combination
(union) of the objects in the lower approximation
and the boundary region. If the boundary region
is empty (i.e. all objects are either in the set
or not in the set), the set is crisp. If the
boundary region is not empty (i.e. some object's
set membership is uncertain), the set is said to
be an approximate (or rough) set.

37
__________________________________ Rough Sets -
Lower approximation ______________________________
____
G In P a m l n b m l n d f m n g f h n e f h y f
f h y c m m y
Concept of no purchase induced by gender and
income LA a, b, d i.e. If G m and In
l, no P If G f and In m, no P
38
__________________________________ Rough Sets -
Upper approximation ______________________________
____
G In P a m l n b m l n d f m n g f h n e f h y f
f h y c m m y
Concept of no purchase induced by gender and
income LA a, b, d i.e. If G m and In
l, no P If G f and In m, no P UA a,
b, d, g, e, f i.e. If G f and In h,
uncertain result

39
__________________________________ Rough Sets -
Lower approximation ______________________________
____
G In P c m m y e f h y f f h y g f h n
a m l n b m l n d f m n
Concept of purchase induced by gender and
income LA c i.e. If G m and In m, P
40
__________________________________ Rough Sets -
Upper approximation ______________________________
____
G In P c m m y e f h y f f h y g f h n
a m l n b m l n d f m n
Concept of purchase induced by gender and
income LA c i.e. If G m and In m,
P UA c, e, f, g i.e. If G f and In
h, uncertain result

41
__________________________________ Rough
Clustering __________________________________
Voges, Pope and Brown (2002) Introduced rough
clustering. All previously published studies had
used rough sets in classification. Extensions of
rough sets theory - Distance measure -
Algorithm for assigning objects to clusters
Voges and Pope (2004) Used evolutionary
algorithm to derive optimum cluster
solution. Cluster solution defined by templates.
42
__________________________________ Rough
Clustering - Templates ___________________________
_______
Any clause of the form D ( a ??? Va ) is called
a descriptor, with the value set Va called the
range of D. A template is a conjunction of
unique descriptors defined over attributes from
B?? A. Any propositional formula T ? a ??B (
a ??? Va?) is called a template of S. The
cluster solution, C, is defined as any
conjunction of k unique templates
C T 1?? , , ? T k?
43
__________________________________ Rough
Clustering - Example _____________________________
_____
Voges and Pope (2004) Evolutionary algorithm
based rough clustering algorithm Applied to a
study of the beer preferences of emerging
drinkers 174 participants were asked to
subjectively rank attributes of beer on a
four-point scale (Very Important, Important,
Unimportant, and Very Unimportant) in terms of
which attributes they considered when making a
purchasing decision. Five attributes image,
packaging, price, alcohol content, and place
sold.
44
Table 1 Rough cluster analysis of beer
preference data
1 Size of upper approximation 2 Size of lower
approximation VI-Very important, I-Important,
U-Unimportant, VU-Very Unimportant, -Dont care
45
__________________________________ Agents ________
__________________________
46
From Objects to Agents Objects Simple concept -
we intuitively operate on the real world in terms
of objects - virtual, computer based objects
correspond to objects in the real world - have
characteristics - can be modelled by various
data structures within the object - can perform
actions on themselves and other objects - use
functions modelled by computer code carrying out
various logical operations on the data
structures Agents Agents exhibit behaviours that
are more characteristic of realistic social
entities - objects with intentions - beliefs,
desires and intentions can be implemented using a
variety of data structures -
interactions between agents can be implemented
via functions within the agent
47
Agent Based Market Simulation Agent simulation
is a computational approach to research that
models the preferences, circumstances,
strategies, interaction, cognitive models and/or
learning strategies of multiple decision-making
agents. Agents Exhibit behaviours that are more
characteristic of realistic social entities. -
memory - reasoning tasks - goal setting,
developing action plans, and causal reasoning -
social behaviour - monitoring other agents,
cooperation, negotiation
48
Modelling dynamic heterogeneous agents reflects
an emerging methodological initiative within
science - Agent Based Social Simulation (ABSS)
in Social Science - Agent-based Computation
Economics (ACE) Applied to studies of
interaction with an environment, group formation,
trade, migration, combat, transmission of
culture, propagation of disease, population
dynamics, etc. Journal of Artificial Societies
and Social Simulation
49
Computer Simulations - software agents
represent components of a business system -
agents' behaviours are programmed via rules to
behave in ways that realistically depict how
business is conducted - often established
through data mining - rules can range from
simple "if-then" clauses to neural networks -
sophisticated machine learning algorithms -
allow agents to learn and modify their behaviour
during the simulation - models are
calibrated against historical data to ensure the
model is accurately replicating the behaviour
of the real system - simulation shows how agents
collective behaviours govern the performance of
the entire system - eg, emergence of a successful
product - powerful strategic tools for
"what-if" scenario analysis - can generate new
strategies for consideration
50
Schelling - Models of Segregation (1969) -
first concerted attempt to apply, in effect,
agent-based computer modelling to social
science - devised a simple spatially distributed
model of the composition of neighbourhoods, in
which agents prefer that a least some fraction of
their neighbours be of their own "type" - found
that even quite low levels of type preferences
produced segregated neighbourhoods - efforts
constrained by the limited computational power
available - only in the last decade advances in
computing have made large- scale modelling
practical

51
Epstein and Axtell - Sugarscape Model (1996) -
modelled group formation, trade, migration,
combat, transmission of culture, propagation of
disease, and population dynamics Marketing
Example - Chang and Harrington (2000) - describe
a computational model of a retail chain as a
multi-agent adaptive system Application
Example - Bios Group - associated with Stuart
Kauffman (1995) - agent-based modelling applied
to a variety of business problems eg. supply
chain management - Proctor and Gamble Application
Example - University of Surrey - 17/18 January -
workshop on Agent-based models of market
dynamics and consumer behaviour - Unilever
52
Conclusion Formal study of computational
intelligence for fifty years Some products
already available - fuzzy logic in washing
machines - voice recognition and synthesis in
phone transactions Commercial benefits will
drive future development
53
Conclusion Will impact on business as an
academic discipline in three main ways domains
of study interaction between internet
marketing and agent-mediated
commerce research methods most CI techniques
are assumption free, unlike traditional
statistics based methods theoretical
explanations the use of agent-based
simulation techniques will allow integration
of micro and macro elements of a market
54
Voges and Pope, 2006 - 22 chapters - 43
authors - 13 countries
55
(No Transcript)