TITLE 9: INTELLIGENCE DECISION SUPPORT SYSTEM

desirespraytownΛογισμικό & κατασκευή λογ/κού

1 Δεκ 2013 (πριν από 4 χρόνια και 1 μήνα)

120 εμφανίσεις




1

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

TITLE 9: INTELLIGENCE DECISION SUPPORT
SYSTEM





Rossilawati Sulaiman


INTRODUCTION


In this chapter we are going to discuss Artificial Intelligence techniques that can be used in
decision
-
m
aking. Our concern is to enhance our decision support system, especially in the
decision making process. We will explore briefly the fields in AI, and discuss more on
artificial neural network and expert system. Some examples are also given on the
applicat
ions that use AI technique in supporting decision
-
making process.

OBJECTIVES

After you have studied this chapter, you will learn:


1.

the fields in Artificial Intelligence

2.

why we use AI technique in decision making

3.

the use of expert system in m
aking decision

4.

the use of artificial neural network technique in decision making

5.

examples of applications using AI techniques in decision
-
making.



2

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM

MIND MAP


Definition

AI DSS










3

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

9.1

ARTIFICIAL INTELLIGE
NCE: CONCEPT AND
DEFINITION


The are many definitions of Artificial Intellig
ence (AI).

Stuart
-
russel:

Tazmania:



Basically, Artificial Intelligence indicates the criteria that a machine has, which can imitate
human thought process. AI can also be referred to the ability of ‘learning’, ‘reasoning’ and
making decision.



Why do we

need AI elements in implementing DSS? Computers are designed to perform a
task from the easiest to the most highly complicated one, based on people needs.
However, computers are not able to learn from experience like human do. Whenever a
decision is going

to be made, computers cannot simply ‘think’ or assess any
consequences out of it. Artificial Intelligence aims to enhance machine behavior so that an
intelligent element can be embedded into the computer during the decision making
process and problem solv
ing. Therefore, the computer will imitate human by taking into
account every factor presented to it. Furthermore with AI added in, computer can then
learn from experience like humans do and this criterion will assist in making a better
decision.




9.1.1

S
ome Artificial Intelligence Fields



This section will briefly describe on some Artificial Intelligence areas. These concepts
usually applied in different domain of problem.


EXPERT SYSTEM

This system uses human expertise stored in the system. The system
will interact with
users to get information and solve the problem. We will go into detail on the system later in
this chapter.



NATURAL LANGUAGE PROCESSING

The goal of the Natural Language Processing (NLP) technology is to allow computer user
to communica
te with computer using languages that they use naturally and finally users
can communicate to the computer like they communicate with another person.



SPEECH RECOGNITION

This technology is the ability of the computer to recognize and understand the spoken

language. The computer must able to understand anyone’s speech and react to voice
command.



4

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM

NEURAL COMPUTING

Neural Computing technology is based on human brain functionality that is transformed
into mathematical model. This model is used in many areas su
ch as in medical, business,
transportation and many more. We will discuss this topic later in this chapter.



GAMES

There are many AI techniques applied in this area. For example the famous Deep Blue, a
chess game application with intelligent feature embed
ded in it.



FUZZY LOGIC

This technique applied to deal with fuzziness of a fact. Fuzzy logic goes further than
Boolean true/false. It will extend the fact to be partially true (or partially false) with certain
degrees.



GENETIC ALGORITHM

This technique
used in searching for pattern from a set of data. This technique is based on
biological evolution of genetic variation.


INTELLIGENT AGENT

This technique used a small program that will be release to a network to do certain task
automatically. A very simple

example is an antivirus program that is automatically detects
any unwanted program in a computer. Other application that is using agent is to do bidding
in an online auction.















Exercise 9.1


1.

Why do we need AI techn
iques in implementing Decision Support System?


2.


Give 5 fields that are currently used in AI.




5

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

9.2

EXPERT SYSTEM

An expert system is a technique in AI that can enh
ance a decision
-
making process.
Basically, an expert, a person with special skill or knowledge will ‘give’ his knowledge to
the system so that the system will ‘think’ like a human when making decision. In this
section, we will discuss about the general arc
hitecture of an expert system, specifically in
knowledge based system. Following that we will discuss about expert system applications
and how they assist human in making decision.




9.2.1

THE GENERAL ARCHITECTURE OF EXPERT SYSTEM:
KNOWLEDGE BASED SYSTEM























Figure 9.1:

Structure of an Expert System


Figure 9.1 shows the architecture of an expert system. Generally, an expert system
contains two major environments; the development environment and consultation
environment, an inference en
gine and the knowledge base.




Development Environment


In the development environment, the knowledge obtained from the expert, together with
other documented information will be transferred to set of rules by knowledge engineers.
The rules will be kept
in the knowledge base.

DEVELOPMENT ENVIRONMENT


CONSULTATION ENVIRONMENT


Knowledge
-
based

Rules/Facts

Inference Engine


Knowledge
Engineer


Documented
Information


Knowledge
Expert


Explanation
Services


Interfaces


User


Sugg
estion
on actions






6

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM



Knowledge Base


Knowledge base will store rules from experts as well as facts related to the domain
problem. Table 9.1 below shows an example of rules describing
ischaemic heart disease

depicted from [arieff,2004]. All facts about
this disease will be collected from heart experts
and then transformed into rules.


Table 9.1: Rules developed to describe
inflammatory heart disease


Rules for

ischaemic heart disease



IF

You have high blood pressure


AND

You have a diabetes


AND

You ha
ve hypercholesterolemia


AND

You feel pain in the chest for a few minutes


AND

You have nausea/vomiting



THEN

Ischaemic heart disease






Inference Engine


An inference engine is used to infer the rules in the knowledge base during the decision
-
making pro
cess. The engine will infer the rules according to the fact given by users and
come out with a conclusion.





Consultation Environment


In the consultation environment experts in the domain
-
specific problem will assist the non
-
experts, or the target user. T
he explanation service will explain in further detail how the
system comes out with a particular solution. The interaction between end
-
user and the
system is done through the interface.




7

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

9.2.2

APPLICATIONS ON EXPERT SYSTEM


There are many expert system ap
plications applied in various kinds of field. Below are
examples on applications that are developed using expert system technique.


R1/XCON


This system was developed by John McDermott et.al. in late 70’s. The system is said to be
the most successful comme
rcial expert systems. It is developed as to help in setting
configuration in VAX computer system at Digital Equipment Corporation. The configuration
is on the basis of customer’s need. The input of the system is the computer characteristics
required by use
rs. The output would be a suggested decision on the computer
specification, so that users can make decisions to buy computer specification according to
their requirements. However, the system has a few shortcomings where the rules have
expanded from time t
o time. New rules will be simply added in the rulebase and after a
while the rulebase become very large. Consequently the system was no longer reliable
and the company needs to rewrite the system to get better result.



MYCIN


MYCIN is another application
in expert system specifically in medical area. It is developed
using LISP programming language by Edward Shortliffe in late 70s. This application helps
in diagnosing infection in blood
-
related diseases. The input of the system is the symptoms
of the specif
ied disease and the output is the diagnosed disease together with the degree
of certainty and the suggested therapy. The system will ask series of questions to the user
as inputs. This application uses more complex questions to get inputs from users. The
i
nference engine uses the rules in the rulebase during the diagnosing process based on
the inputs given by the user.



E
-
PADDY


E
-
PADDY is a diagnosis and advisory system developed by a group of researchers in
National University of Malaysia [ROSS et. al].

The system covers paddy diseases and pest
control diagnosis on paddy plant. Up to now, the system is able to diagnose 18 possible
diseases, using production rules and frame approach. The system will request input from
the user, to get symptoms of a specif
ic disease by presenting series of questions to the
user. The input given will be matched to the set of rules in the knowledge
-
base to produce
the specific output regarding the disease.










Exercise 9.2


1.

What is the role of inference engine in an ex
pert system?


2.


Give two examples of applications that use expert system technique and
describe how they help in making decision



8

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM

9.3

MACHINE LEARNING

Another technique in AI that can be used

in DSS is Machine Learning. This technique
allows machine to
learn

or acquire knowledge from experience (such as historical data).
Basically, there are two types of categories: supervised and unsupervised learning.
Supervised learning indicates a process
of learning that induces knowledge from a set of
data, which the final outcomes are known. For example, we induce a set of rules from
historical paddy diseases data. We already know all possible cases of diseases that might
occur. Meanwhile, unsupervised l
earning is used to obtained knowledge from a set of
data, which the final outcomes are unknown. For example, we can do classification on
users’ preferences of different products. We do not know what kind of choices the user
might select. Figure 9.*** below

indicates a taxonomy of learning machine, depicted from
[efram Turban]

























Figure 9.2:

Taxonomy of Learning Methods



From Figure 9.2, we can identify several methods and algorithms used in machine
learning:

Machine

Learning

Supervised


Unsupervised

Explanation
-
based learning

Statistical
regression

Inductive
learning

Case
-
based
reasoning

Neural
Network

Genetic
Algorithm

Genetic
Alg
orithm

Neural
Network

Clustering




9

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM




Inductive Learning


Indu
ctive learning involves the process of learning by example. The system will do a
generalization from a specific observation. The generalization can be used for
explanatory or predictive purposes. For example in explanatory purpose, the system
will produce
a set of similarities from a given set of data. For instance, a customer
buying
-
behavior study discovers that if customers buy sausage and bread, it is likely
that they will also buy butter. In predictive purpose, the system will learn from a set of
data t
hat is classified into two or more classes. For example, the system is required to
classify different types of animals.





Case
-
based reasoning


A case
-
based reasoning (CBR) system works by finding solution to new problem from
a historical database and then

adapting successful solutions from the past to current
situations. For example, if a computer does not start, CBR can be used to match the
characteristics of the problem with a database.




Neural computing


Neural computing is a technique that attempts to
imitate the structure and functionality
of the human brain. The system uses this technique will be presented to a set of
training data so that the system can learn to solve a specific problem. More on this will
be further explained in section**.





Genetic
algorithm

Inspired by Darwin's theory of evolution, this technique solves problem by using
evolutionary process. Initially the system is presented by a set of solutions that is
represented by chromosomes. The set is also called population. The system will
try to
find the right solution from the population. The chosen solutions will be used to form a
new population, which is possibly better than the previous one. The selected solutions
are based on their fitness. This process is evolving until some condition

is satisfied.




Clustering

put data into several groups based on their similarities, used for
marketing etc



Statistical methods
-
such as multiple discriminant analysis, suitable for analyzing
knowledge that is quantitative in nature and have been applied
to knowledge
acquisition, forecasting, and problem solving



Explanation
-
based learning
-
combine exixting theories to explain why one instance
is not a prototypical member of a class








10

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM












We have discussed so far that there are many methods and alg
orithms used under
machine learning that can be an aid to decision support. In the next section, we will be
discussing about neural network method in detail and how it can be used to support
decision
-
making.
























11

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

9.4

ARTIFICIAL NEURAL NE
TWORKS
(ANN)

The real human brain biological process initially inspires Artificial Neural Networks (ANN).

ANN is one of the AI techniques that can support decision
-
making. ANN can learn from
experience like human do. Generally, neural networks are used in pattern

recognition,
generalization, and prediction. Neural networks are trained by showing them examples
data repeatedly. The data contain the input and the desired output for a particular problem.
The network will be trained to learn to achieve the desired outp
ut. When the network has
succeeded in mastering the learning process, we can test the network by a new test data
that the network never seen before. At this stage, the network should still be able to predict
the correct output.


Neural network software p
ackages are widely used in business especially in stock market
prediction, credit card fault detector and many more. In medical area, neural networks are
widely used in health monitoring and diagnosing diseases. You can browse
http://www.calsci.com/Applications.html

for more examples of neural network
applications.



9.4.1

Basic component of ANN


ANN consists of several interconnected simple processors as a parallel computing
procedure. Each processor

(also can be referred to as a node or neuron) is only concern
on signals it sends to and signals it receives from other processors periodically. Each
processor will be cooperating with other processors in a large network to perform the
required task. For
example, in business area neural network can be used to help us decide
whether a candidate is qualified for a bank loan. There are also applications that can help
us forecast the stock market prices.





The processing unit

Now let us take a look at the bas
ic component of ANN. Figure 9.1 below shows a single
neuron in a network. The neuron will receive several signals as input from other connected
neurons as well as sending its signal to the other neurons in the network as an output. The
output signals sent
to the other connected units are also known as
weights
. Each neuron
will have the capability to compute the combining input signals. An activation rule in each
unit will later compute the output signal using the combining input value. The output will be
th
e input to other neurons, but it might as well be the final result such as 1 as YES and 0
as NO.











Net j=


Wij . xi


Input signals (weights)
from other processors

Output signals
(weights) to
other processors

X
1

X
2

Xi

Y
j

W1
j

W2
j

Wi
j

ƒ
(x
)

Activation
function

Neuron
j



12

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM


Figure 9.3:

A single processor in a network

From figure 9.1,
x1, x2,…, xi

are the values of input neurons, with weights values of
w1j,

w2j,…, wij

r
espectively. The calculation of the combined input
Netj,
is the summation of
multiplication of input value
xi
with the respective weights
wij.


Netj

=


n
i
xi
Wij
1
.



where n is the number of connected neurons,
Wij

is the weight sent to the neur
on
j

and
xi

is the values of input neuron. There is also another function to calculate the output signal,
called the activation function,
f(x).

This function is actually to normalized the value of
combined input value,
Netj

that later be the final output v
alue for the neuron. This output
value will either be sent to the other neurons as input value or it might be the final output
for the network.



EXAMPLE 9.1


This example will show the calculation of combined input
Netj

for neuron
j
illustrated in
figure
9.**.




















or alternatively, we can use matrices like the following:




2
.
0
4
.
0
5
.
0










3
.
0
4
.
2
2
.
0

= 1.12




Figure 9.4:

An example of calculating the combining output in a ne
uron

j

0.2

2.4

0.3

0.4

0.2

0.5

Netj

=



n
i
xi
Wij
1
.






= (0.5 X 0.2) + (0.4 X 2.4) + (0.2 X 0.3)






= 1.12





13

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM




14

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM

















The network

One of the main aspects to be considered in ANN is how to construct a network, which
requires us to concern on the connectivity among neurons, including the direction of the
connections as well as their respective weight v
alues. Each ANN consists of neurons
arranged in layers. Basically, there can be three layers of neurons illustrate in figure 9.***
below. There are input layer with four neurons as input units, middle layer with two
neurons, which is also known as hidden l
ayer, and output layer with one neuron as output
unit. It is called input unit because it will take input directly from the environment like
keyboards, sensors and so on. The same thing applied to output unit, which will send
output directly to the environ
ment. Meanwhile, middle layer or the hidden layer consists of
hidden units and they are not directly connected to the environment.


Each neuron is connected to the other neurons in the adjacent layer. There can be more
than one hidden layer depends on how

we model our network. The neurons can be
connected to one another in various ways. Each neuron can be connected to every other
neuron, or each neuron can only be connected to the adjacent neuron on the other layer.
Also, there are models that allow feedba
ck connection to the other neuron on the adjacent
layer and so on.




















Output lay
er

Hidden layer

Input layer

Exercise 9.3


Calculate the combined input for the following neuron:




Neuron j

W1=0.2

W2=0.4

W3=0.7

X2 = 5

X3 = 1

X1 = 3




15

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM



Figure 9.5:

A Neural Network with single hidden layer





16

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM



Activation Functions

An activation function in a network will carry out a task of calculating the output signa
l of a
neuron. The calculated output signal then sent to the other neurons. The final output could
be a real number, or real number within a tolerance interval [0
-
1], or a discrete number like
{0,1} or {
-
1,+1}. A simple example of activation function is th
e Sigmoid function, which is
commonly used in ANN. The output from this function is in the range of 0 to 1. An example
is shown in Figure 9.*** below. The function is involve is:


)
exp(
1
1
Netj
Yj



.


Yj
is the normalized value of combined input value

Netj,
typically between 0 and 1. This
process is performed before the output sent to the other neuron in the next layer. If the
normalization is not done, the output will be very large and it is difficult to reached the
desired output of the network.













Figure 9.**:

The sigmoid function


Apart from the sigmoid function, there is also function that uses a treshold value such as in
Binary treshold function. This function will limit the
Yj
value to 0 or 1 depending on the
value of
Netj

relative to
a particular treshold value,

. This is illustrated in Figure 9.***
below.













Figure 9.6:

The Binary treshold function


The process in the activation function can take place at the output on each neuron, or
done only at the

output layer, depending on how we model our network. Next, we will see
an example of a network that learns the exclusive
-
or (XOR) arithmetic.

0.2

0.4

0.6

1.0

0.8

-
4

-
2

2

4

Netj

Yj

1.0

Netj

Yj


1

if
Netj









if
Net
j <


Yj =




17

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

EXAMPLE 9.2


This example will show us a network that learns the XOR relationship. Table 9.1 shows the
XOR opera
tor's truth table:


Table 9.1: XOR relationship


INPUT

DESIRED
OUTPUT

X1

X2

1

1

0

1

0

1

0

1

1

0

0

0


The output can only be true if either x1 or x2 is true (1). The network is designed to be a
layered feedfoward network, with 2 input units, one hidd
en layer with 2 hidden units and an
output unit. Feedforward means that the connection of each neuron will only be in one
direction.













Exercise *** CE BUAT LATIHAN PENGIRAAN



1.

Gi
ve three examples of Groupware that support GDSS and describe their

criteria briefly.




18

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM



A learning rule to adapt the weights


There are many learning algorithms in ANN used according to the require
d tasks. There
are two groups of learning algorithm in ANN; supervised and unsupervised learning
illustrated in Figure 9.***, depicted from [L. Medsker and J. Liebowitz, 1994].
































Figure 9.**:

Learning Algorithms And Architect
ure In Neural Network



As discussed before, supervised learning involves a learning process from a given set of
data of which the final outcomes are known. For example a set of historical data on
diagnosing heart attack is presented to a network as an inp
ut. The desired output should
be equal to 1 if the patient had a heart attack and 0 if the patient is healthy. The network
will work on the data and produce the actual output. It will learn to produce output as
closed as the desired output by calculating t
he difference between the two outputs.
Another version of this technique is to consider the previous weight change when
attempting to do error correction as in the backpropagation and the Hopfield network.


Meanwhile, in unsupervised learning, the network

is left to learn on its own with no desired
output (self
-
organized). The network will cluster particular patterns of output that have
something in common. For example, the network is trained to recognize jockey players
Learning Algorithm

Discrete/Binary input

Continuous input

Supervised

Unsupervised

Simple Hopfield

Outerproduct AM

Hamming n
et

ART
-
1

Carpenter/
Grossberg

Delta rule

Gradient descent

Competitive
learning

Neocognitron

Preceptor

ART
-
3

SOFM

Clustering
algorithm

Supervised

Unsupervised

Nonlinear vs
Linear

Backpropagation

ML perceptron

Boltzman

Hopfield

SOFM

ART
-
1ART
-
2

Architecture

Supervised

Unsupervised

Recurrent

Feedforward

Estimators

Extractors




19

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

and sumo players, which obviously us
e the weight parameter in clustering. Therefore, the
network must able to group data that shared relatively similar weights. This kind of self
-
learning can be seen in Adaptive Resonance Theory (ART) and Self
-
organizing Feature
Maps (SOFM).


Generally, supe
rvised learning process in ANN involves processes of:

a.

Computing the output from input layers,
Yj

b.

Comparing
Yj

with the desired output,
Z
.

c.

Adjusting the weight and repeat (a) and (b) to minimize difference

,
between Yj
and Z


Contoh back propagation

examp
le




20

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM






21

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

The main goal of learning process is to decrease the value of


so that
Yj

is as closed to
Z

as possible by adjusting the weights.



9.4.2

Example


ANN has been used in many applications such as in credit card





Now we will discuss examples on t
he usage ANN in decision support. We first start with a
very simple example that is to calculate





22

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM

9.4.3

Example of calculation












Exercise 3.4



1.

Giv





23

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

9.5

OTHER TECHNIQUES IN
AI



























SUMMARY

GDSS concept, supported by its technologi
es can be applied in group decision
-
making.
GDSS technology is meant to enhance group performance and decision
-
making and less
time consuming. There are four type of collaboration that helps in decision
-
making. The
decision process can be in a meeting roo
m, where all members can meet at the same
time. It can also occur among geographically distributed members. This can be done by
using various GDSS Groupware. GDSS concept can be applied not only in the
organization, but also in education.




TEST 1

Inst
ruction: Answer all questions in exactly 15 minutes.


1.

Describe four types of collaboration.

2.

Give 3 examples of group supporting tools.

3.

When would you use various group support tools?


Exercise 3.5



1.

E
xplain how GDSS concept c
an be applied in education sector.




24

OUM


TITLE 9


INTELLIGENCE DECISION SUPPO
RT SYSTEM

4.

How can videoconferencing and groupware helps in an organization that
has few
branches?

5.

What is a definition for a "Virtual Organization"?





25

OUM


TITLE 9


INTELLIGENCE DECISION SUPPORT SYSTEM

TEST 2

Instruction:
Answer all questions in exactly 30 minutes.


1.

In what condition do managers need the support of GDSS?

2.

What advantage can managers get from the support p
rovided by GDSS?

3.

What type of technologies used in order to make a “Virtual Workplace” possible?

4.

What is a definition for a "Virtual Organization"?

5.

What is the advantage and disadvantage working in Virtual Organization?






Reference


L. Medsker and J.
Liebowitz, Design and Development of Expert System and Neural
Computing, New York: Macmillan, 1994.