complete_project
TRANSCRIPT
ABSTRACT
The purpose of this research is to determine the performance of a
Logistics Network using the techniques of Artificial Neural Network. By
analyzing the performance and comparing the predicted values to the
industry standards of the company, the success of a Logistics Network
can be predicted. In the first segment we work with a sample Network
matrix and find out the shortest cost-distance path. Then we try to find
any similarity for relationship between them. In the second segment we
work with a more complicated involving large number of parameters
affecting the Logistics Network and try to find the relationships. Once the
network is initialized, prediction is made which is then compared to the
industry data & tradeoffs. The computing is done with MATLAB and
various plots are generated. We also looked in the theoretical aspects
involving this project research while studying neural networks in detail.
The project completed with a sense that given sufficient historical data
neural networks can be used to make intelligent predictions for various
purposes including Logistics Network.
KEYWORDS
Logistics Network
Artificial Neural Network
Artificial Intelligence
Business Predictions
MATLAB
Supply Chain Management
Feed-forward Back-propagation Algorithm
PREFACE
This project report is being formulated and prepared by me on the basis
of my two year Master in Business Administration ( MBA ) degree course
in Indian Institute of Engineering Science & Technology, Shibpur during
the term 2014 - 2016. This project has been done in requirement to the
partial fulfilment of the Master’s degree. During the period, I had 24
papers with 6 papers per semester in addition to the SIP, Assignments &
other projects. This entire research work is prepared by me with the help
of the experience & knowledge gathered during the course duration.
Besides enhancing the subjective knowledge, the MBA course had helped
me to make out the rules and regulations of the business world, the
industry standards followed by the experts and techniques to manipulate
& run the business world. All these have made the course worthy.
As I am having a technical background of Electronics & Communication
Engineering, I wanted to do my project where I could apply by technical
knowledge as well as implement by newly learnt management
methodologies. Hence, I decided on the topic “ Predicting Success of a
Logistics Network using Artificial Neural Network “ which looks into the
ways we could predict a future outcome based on historical data from
past businesses. This project has been very instrumental in sharpening
my management knowledge and made me feel refreshed with
information and new ideas upon completion.
..............................................................
Date: ...................................................
ANIRBAN ROY
School of Management Sciences
Indian Institute of Engineering Science & Technology
ACKNOWLEDGEMENT
I, the undersigned, student of School of Management Science ( SOMS ) of
Indian Institute of Engineering Science & Technology, Shibpur would like
to thank the institute for giving me the opportunity to study in this
prestigious university.
Besides my parents, I would like to thank my teachers for giving me all the
knowledge, advices & instructions for these two years. I believe that the
knowledge and professional expertise I acquired in these two years would
help me rest of my life to lead a better life. I am indebted to them.
Next, I would also extend my sincere gratitude to library madam,
computer madam and supporting staffs for giving me all the support
necessary for the course duration. Without them, it would not have been
so easy.
Next, I would like to thank our Director, Mr. Ajoy Kumar Ray & SOMS
Director Mr. P.K. Paul for supporting us whenever possible and helping us
achieve best possible education.
Finally, I would like to give sincere respect to all the authors, researchers
& scholars whose work helped me in completing my project work. I would
like to thank Wikipedia for sharing us the vast resource of knowledge and
the creators of MATLAB for giving the world a powerful tool which
simplified my tasks to a great extent.
..............................................................
Date: ...................................................
ANIRBAN ROY
School of Management Sciences
Indian Institute of Engineering Science & Technology
CONTENTS
Title / Subject Matter Page
1. INTRODUCTION ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... .... 1
2. ARTIFICIAL NEURAL NETWORK ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... . 1
a. Learning Rule 3
b. Neural Network Model 4
c. Neural Networks in Business 6
d. Neural v/s Logistics Network 7
3. LOGISTICS NETWORK ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... . 8
a. Types of Logistics 8
b. Nodes of Logistics Network 8
c. Network Optimization 9
d. Network Optimization Algorithm 9
e. Conclusion 9
4. RESEARCH METHODOLOGY ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... 10
5. RESEARCH IMPLEMENTATION ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... 11
• Shortest Path Rule 12
• Sample Example 1 13
• Example 1 Coding 15
• Example 1 Neural Network 16
• Sample Example 2 21
• Example 2 Neural Network 23
6. CONCLUSION ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... 25
7. RECOMMENDATIONS ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... 26
8. REFERENCES ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... . 27
9. APPENDIX ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... . 28
1 | P a g e
1 Predicting Success of Logistics Network using Artificial Neural Network
1. INTRODUCTION
In this era of 21st century every business
entity would want to know if their business
plan would succeed in the future. And, the
survival of the company can depend upon
the business plan executed. Not only so,
the quality of execution, the time of
delivery, nature of strategic steps and also
the post plan execution steps are also
important. Logistics and supply chain
management is an integral part of any
business organization. Logistics, in short is
all the operations in the process of flow of
materials from the suppliers to the
customers or the point of origin to the
point of consumption. A good logistics
network ensures that the required
resources are available at the required
place at highest possible quality and at
lowest possible cost. Investors and
businesses alike use several methods and
tools to evaluate a business plan to
measure its future impact even before its
execution.
Integrating a model that takes into
account both the quantitative and
qualitative regression model is quite
tremendous task as there are a lot of
variables that could affect the business
process or the business as a whole. There
is also a limitation to ANN which is that
this model needs sufficient base or
historical data to make appropriate
computations.
In corporate business world, Neural
Networks can be used to study about
stock markets (predictions of stock market),
marketing of a product (acceptance of a
product by customer in a target market,
longevity of a product, product development
cycle), performance appraisal, financial
assessments, etc. The list goes on and on.
In short it can be inferred that ANN model
can be used to solve practically any sort of
complex problems in all the fields of
management sciences starting from
Finance to Human Resource and
Marketing. It can also be employed at
production sites and assembly lines to
increase product quality & reduce wastage
of time and resource. Hence, fairly good
predictions can be made by managers
based on it.
Our objective is to study a logistics
network & analyze it using algorithms to
find shortest cost path and the shortest
distance. We also make predictive
assessment of a multi-variable logistics
network, which is later used to determine
its success or failure. Finally, based on the
analysis it can be understood if the
network needs to be optimized. But, the
study can be further done to discover the
weak links in a network & provide a
mechanism through which automated
adaptation to logistics network takes
place. MATLAB is used for programming
and generating the processed data. The
project work ends with stating the
limitations and illuminating the future
study which can be done.
2. ARTIFICIAL NEURAL NETWORK
To study in detail how is ANN used to
predict possible outcomes, we must first
know perfectly well about ANN. Wikipedia
defines Artificial Neural Network (ANN) as
follows:-
Artificial Neural Network is a model used to
estimate or approximate functions that can
depend on a large number of inputs and are
generally unknown. Artificial neural networks
are generally presented as systems of
interconnected "neurons" which exchange
messages between each other. The connections
have numeric weights that can be tuned based
on experience, making neural nets adaptive to
inputs and capable of learning.
Artificial Neural Network is basically a
model which predicts an outcome based
2 | P a g e
2 Predicting Success of Logistics Network using Artificial Neural Network
on permutations, combination and
probability. And, with the presence of large
volume of data (which acts as a random
data having high credibility) near perfect
assumptions can be made using this
model. ANN is a continuous learning
model which means that this model
continuously evolves with time thereby
decreasing the deviations.
In terms of logic, ANN consists of
interconnected nodes. Neural Networks
got its name from the biological neural
networks of brains. Just like the neural
chain, ANN consists of node points which
are connected together to form an input-
output relationship. In case of complex
ANN there may be some transient node
which works as interlink between input
and output nodes.
Fig 1. Artificial Neural Network Model
This transient node is hidden from both
the input and output point of view and it
works as a processing or decision making
node. Its called “Artificial” Neural Network
as it has Artificial Intelligence based on
coding as it is capable of altering itself
based on the inputs and situation
presented. As a result it is adaptive in
nature. A neural network is made up of
simple calculating units called Neurons
that locates in different layers and have a
great many internal connections.
For several decades, forecasting
and prediction has been largely made by
linear simple approaches which are very
simple to understand and implement. But,
these linear methods fail to grasp the non-
linearity of complex variables and their
limitations as a variable. There can be
different cases with respect to variables
such as:
• Some variables can affect a system
more than other variables. These
variables are called mission critical
variables.
• Some variables have negligible or
no effect on a system.
• Sometimes variables remain
constant throughout a cycle while at
others they are random.
A wide plethora of problems has been
successfully solved by applying Artificial
Neural Networks. ANN mostly consists of a
learning rule which governs the process
and enhances the connections between
the nodes based on a pattern. Just like
human brain learns from experiences;
ANNs do the same. Thete are two major
life-cycle of ANN:
� Training Cycle
The process is implemented in ANN &
is fed with random data. The network
learns from their experience which is
used to enhance the connections.
During this period, output is only error
data which is fed back into system to
make it more accurate.
� Run Cycle
Actual running of the system takes
place. Output acquired is of good
quality but even now there is presence
of feedback mechanism to make the
system more robust to variations.
3 | P a g e
3 Predicting Success of Logistics Network using Artificial Neural Network
2 (a). Learning Rule
There can be different types of learning
rules of Artificial Neural Network which are
used to teach the network to make near
perfect decisions are given below.
� Unsupervised Learning
In case of Unsupervised learning the
system is fed with raw data so that it could
adjust itself with variations. But, this
system of learning lacks second set of data
which can be compared with the outputs
received. So, errors in the system cannot
be detected. The system has to use the
concepts of probability density function to
replicate what exact value would be in the
presence of sufficient amount of data.
Fig 2. Probability Density Function
Probability Density Function tell us about
the relative likelihood for a random
variable to take on a given value. The
standard normal distribution has
probability density
���� = �√� ���/�
When X (random variable) is given and it
has a probability density function f, then
the value of X can be calculated as follows:
��� = � ����������
There are also various approaches to
unsupervised learning such as blind signal
separation, clustering, method of
moments, etc; discussing them is out of
the scope of this article.
� Supervised learning
It is the process of learning from
supervised or labelled data which are
otherwise known as training example.
Each training example consists of a input
vector and estimated output vector. When
this key-value pair is fed into the neural
network and output is compared.
According to the variation in the output
result, the processing function or the
algorithm is modified which can be used to
analyze further examples. A good example
of supervised learning in case of human is
concept learning, where teachers provide
the required input vector. Whenever
student deviates from the logic, the
teacher makes necessary corrections. By
following this way repeatedly, the student
learns the concept or the logic.
� Reinforcement learning
In case of Reinforcement learning, a
input/output pair is provided to the
system. But, the difference between
Reinforcement learning and supervised
learning is that the prior never uses the
correct input/output pair nor the correct
response is given any special care. The
system is based on reward system and the
learning agent interacts with the
environment in discrete time duration.
During these discrete time steps, input is
provided to the system and observation is
made. Reward is also given to
acknowledge the observation. The basic
goal of Reinforcement learning is to collect
as many rewards as possible.
4 | P a g e
4 Predicting Success of Logistics Network using Artificial Neural Network
After running for sufficient amount
of time frame, the past performance is
compared to the present performance.
Hence, comes the concept of regret which
is used to make necessary adjustments to
the system. The reinforcement learning is
used in case of long-term decision versus
short-term decision trade-off. This system
can be used in the following situations:
� No analytic solution is available
which can be applied to the context
� Only the model of the situation is
known
� Only way to get information about
the environment is by simulation or
interacting with the environment
directly
� Need to evaluate effect of long-
term/short-term decisions
2 (b). Neural Network Model
A Neural Network is basically an inter-
connection of layers of nodes. Each
individual connection has some weight
and when input is provided to it, an output
is passed to the next node based on the
connection known as synapse. These
“weights” are the basis of manipulation of
input. During learning mode, these
weights are regularly changed or updated
to match the required output.
Fig 3. Schematic of a Neuron
In the above figure, x1, x2, .... xn are the
inputs from which signal is entering the
neuron. And, each of these inputs carry
respective synaptic weights of w1, w2, .....
wn. Hence, there are n numbers of total
inputs given to the network. Hence, the
total weighted input from all the inputs is:
� = � �����
���
ξ is the excitation level of the neuron,
gives the output y after reaching excitation
level h, which is the maximum excitation
level of the neuron. The level of excitation
h is achieved by activation function σ (sigma).
���� = 1 "# � ≥ ℎ 0 "# � < ℎ (
The above sample function can also be
called as limiting function as it provides
output only two discrete units; one or
zero. So, the complete mathematical
formulation of the artificial neuron is given
by the following expression:
���� = 1 "# � ≥ ℎ 0 "# � < ℎ ( �ℎ)*) � = � �"�"+
"=0
Signal propagation through the neural
network and its manipulation can be done
by either changing either the weights of
the synapse connections or the path
followed by the signal itself. The grouping
of the neurons, synaptic weights and the
activation potential is together known as
the configuration of the network. Other
examples of activation function could be
as follows:
� piece-wise linear function
���� = ,1, � > 1�, 0 ≤ � ≤ 10, � < 0(
5 | P a g e
5 Predicting Success of Logistics Network using Artificial Neural Network
� hyperbolic function
���� = 1 − )�11 + )�1
� sigmoid function
���� = 11 + )�1
There are two states in which the neuron
could stay. Either the electron is in firing
state or it is in state of rest. The two states
of individual neuron can also be visualized
as a vector containing both magnitude and
direction. The weight can be taken as the
magnitude and the state of electron can
be taken as the direction.
Take an example to clearly
understand the above concept.
x1 w1
c1
x2 w2
c1=x1w1 + x2w2
The output electron is fired only when the
value of c1 is greater than the threshold
value of the network. In the same way
larger complex network can be realized.
Now for the sake of example, let the
Neural Network mimic a ‘AND’ gate. A
hypothetical unknown logic is provided to
the neuron. The truth table of a AND gate
is as follows:
x1 x2 c1
0 0 0
0 1 0
1 0 0
1 1 1
A threshold value of q is required so that
the electron is fired. Imagine in the Neural
Network, x1 and x2 are provided with
certain weight and then random states of
electron are observed. Suppose the
desired output of the network is d and the
real output of the network is y. If the Then
the error e of the network would be as
follows:
e = d – y
Then error function of the network in the
long run would be as follow:
= �� ��� − 3��
The goal of the network is to minimize the
error to the maximum extent possible. In
order to do so, the weights of the network
are regularly updated after each cycle till
the error tends to be zero. Suppose in the
sample example above, both w1 and w2 are
random variable. The network is allowed
to run without any learning process. The
following table is acquired:
x1 |x1| x2 |x2| y |y| d e
0.61 1 0.85 1 0.34 0 1 �
0.84 1 0.59 1 0.95 1 1 --
0.20 0 0.09 0 0.56 1 0 �
0.57 1 0.83 1 0.22 0 1 �
0.12 0 0.31 0 0.69 1 0 �
0.29 0 0.60 1 0.59 1 0 �
0.55 1 0.80 1 0.96 1 1 --
0.34 0 0.53 1 0.30 0 0 --
0.50 1 0.48 0 0.10 0 0 --
0.34 0 0.51 1 0.25 0 0 --
0.50 1 0.63 1 0.34 0 1 �
0.34 0 0.26 0 0.49 0 0 --
0.42 0 0.97 1 0.09 0 0 --
0.43 0 0.79 1 0.56 1 0 �
0.26 0 0.54 1 0.84 1 0 �
0.77 1 0.85 1 0.96 1 1 --
0.76 1 0.38 0 0.37 0 0 --
0.88 1 0.11 0 0.63 1 0 �
0.24 0 0.51 1 0.79 1 0 �
� indicates presence of error
Here the sample size, n = 20. Then,
4 = ����� − 3��
�5
6��
4 = �� 7�58
6 | P a g e
6 Predicting Success of Logistics Network using Artificial Neural Network
Hence, Z = 5. And, the probability of
occurrence of error in this case would be
9�:� = ;<=>?@<. <#BCCD*>+C)<#:**<*EF>GH?)F"I)
J� � = �5�5 = 5. K5 = K5%
Hence, from the above data, it can be
inferred that there is 50% chances of
getting incorrect solution. The above
example is based on simple random
weights and it lacked a learning
mechanism which could train the network
to make better choices with the passing
time.
The above example can be modified
to work as Artificial Neural Network by
training the network to rectify its strategy
whenever it receives incorrect answer
during the training exercise. We assume
that each sample is obtained from each
instance of time period. We need to obtain
a situation where there is output value,
which we know is possible only when the
output is greater than the threshold value.
Suppose, in the n � 0 instance of
time, we get output across the points as:
x1 |x1| x2 |x2| y |y| d e
0.02 0 0.13 0 0.64 1 0 �
Here the desired output is 0 but we get an
error. So, the learning mechanism teaches
the network in that instance d � 1 for
x1=0.02 and x2=0.13.
In the next round, n � 1 instance of
time, we get another output as follows:
x1 |x1| x2 |x2| y |y| d e
0.25 0 0.36 0 0.52 1 0 �
Here also, the desired output is 0 but we
get an error. So, the learns that in that
instance d � 1 for x1=0.25 and x2=0.36. In
the same way it also learns about the
other logic states of ‘AND’ gate and finally
it learns to mimic the truth table. Slowly
the network learns to accept values from
0.50 and above as logical 1 and values
from 0.49 and below as logical 0. It also
learns about what the output values
should be for respective values of input.
The computer modelling of this sample
example is out of bounds of our work. But,
we can understand from this example that
a more complex network works in the
same principle. In the same way it learns
and this auto learning can be referred to
as Artificial Intelligence in computer
terminology.
2 (c). Neural Networks in Business
To fully understand the practical use of
Artificial Neural Network in the field of
business, we must first keep in mind that
the practical simulation of a business
model may not exactly look like a network
of neurons. But, it follows the same
concept of giving weighted inputs to a
network; where network can be seen as an
algorithm or conditions or limitations, and
getting an appropriate result to satisfy the
situation.
In this piece of work, we would look
upon various models and theories which
are actively used to make real-life
decisions in a business environment.
Some examples of Neural Networks in
Business environment would be as
follows:
� Predicting stock market simulations
� Simulate introduction of a new
product in the market
� Properly analyze seasonal
variations
� Analyze customer related data for
making predictions and finding out
similarities
� Project management & formulating
strategies
� Financial and economic forecasting
7 Predicting Success of Logistics Network using
� Solving transportation problem and
simplifying Supply Chain
� Tourism Demand Forecasting in
Tourism Industry
� To predict bankruptcy of a business
� Inventory management to reduce
both inventory levels
� Risk assessment & prediction
� To analyze proper investments of
shares and funds
� Predict and analyze the efficiency of
a logistics or supply chain network
including its cost benefit
The science of Artificial Neural Network is
evolving and a lot of work in this field has
been experimental. As these Neural
Networks can get really a lot complicated
with each addition of nodes and logic used
in the neurons, these neural networks can
only be tested meaningfully using software
codes and logics. Software that has been
used a lot in this field is MATLAB
saw very well that the practical use of
Neural Network is endless in real business
environment. We take an example of
predicting business success by Neural
Networks using MATLAB
algorithms.
2 (d). Neural v/s Logistics Network
There are lot of similarities to a
Network and a Logistics Network
the network is a multi-nodal network. Just
like a Neural Network which consists of
numerous hidden layer nodes, so does the
Logistics Network. For example, the raw
material which is acquired from the
suppliers travels through various means of
transportation to reach the industry and
from there after the product is made again
it travels through various means of
transportation to reach the end users.
And, there are a lot of factors influencing
these nodal points like storage capacity,
Predicting Success of Logistics Network using Artificial Neural Network
ransportation problem and
Supply Chain
Tourism Demand Forecasting in
To predict bankruptcy of a business
Inventory management to reduce
Risk assessment & prediction
To analyze proper investments of
Predict and analyze the efficiency of
a logistics or supply chain network
including its cost benefit
The science of Artificial Neural Network is
evolving and a lot of work in this field has
been experimental. As these Neural
Networks can get really a lot complicated
with each addition of nodes and logic used
in the neurons, these neural networks can
be tested meaningfully using software
codes and logics. Software that has been
MATLAB. So, we
very well that the practical use of
real business
We take an example of
ess success by Neural
MATLAB software
Neural v/s Logistics Network
There are lot of similarities to a Neural
Logistics Network as both of
nodal network. Just
like a Neural Network which consists of
numerous hidden layer nodes, so does the
Logistics Network. For example, the raw
material which is acquired from the
suppliers travels through various means of
ransportation to reach the industry and
from there after the product is made again
it travels through various means of
transportation to reach the end users.
And, there are a lot of factors influencing
these nodal points like storage capacity,
transit duration, cost involved in storage
and movement, etc. The main nodal points
of a logistics network can be abbreviated
as follow:
Fig 4. Abbreviated Logistics / Supply Chain
Network
The Logistics network in terms of nodal
points can be visualized as below
hidden layer consisting of warehouses,
retailers, stockers, etc. And, we see that on
the whole it is a N:N relationship matrix
just like a Neural Network.
Fig 5. Nodal points of a Logistics / Supply Chain
Network
Supplier
Retailer
Customer
7 | P a g e
tion, cost involved in storage
and movement, etc. The main nodal points
of a logistics network can be abbreviated
Abbreviated Logistics / Supply Chain
Network
The Logistics network in terms of nodal
points can be visualized as below with the
hidden layer consisting of warehouses,
retailers, stockers, etc. And, we see that on
the whole it is a N:N relationship matrix
just like a Neural Network.
Nodal points of a Logistics / Supply Chain
Network
Manufacturer
Distributor
Customer
8 | P a g e
8 Predicting Success of Logistics Network using Artificial Neural Network
3. LOGISTICS NETWORK
To start with this topic, our first and
foremost goal must be to understand the
two terms: Logistics & Supply Chain
Management. Both of these terms are
sometimes used interchangeably. But are
these terms really mean the same? It’s
rather confusing. According to Logistics &
Supply Chain Management by Martin
Christopher:
Logistics is the process of managing the
procurement, movement and storage of
materials, parts and finished inventory and its
marketing channels so profitability are
maximised. Supply Chain Management is the
management of upstream and downstream
relationships with suppliers and customers in
order to deliver superior customer value at less
cost.
The following two terms can further
be simplified by stating exactly what
comes under which term. The components
of Logistics are:
• Customer Service
• Purchasing & Procurement
• Warehouse Management
• Transport
• Production Planning
Likewise, the components of Supply Chain
Management include the following:
• Suppliers
• Manufacturer
• Retailer
• Stocker & Wholesaler
So, Logistics is the smaller domain while
Supply Chain Management is the larger
domain. Supply Chain Management is the
integrating factor of Logistics as it
comprises the Logistics part as well as the
management of all supporting
departments to bring competitive
advantage to a business organization. In
terms of mathematics, we can say that
Logistics is a subset of the bigger SCM.
The main aim of Logistics Management is
customer satisfaction by providing
efficient transportation of resources,
people, information or energy between
two points. A robust Logistics network
helps in timely and cheap transportation
facilities. And; in today’s highly globalized
industry, quick and timely delivery can be
the basis of survival of a company.
3 (a). Types of Logistics
There are basically two types of Logistics
segment which can encompass all
Logistics activities. The two major
segments are:
(i) Inbound Logistics
They describe all logistics activity which
focuses on purchasing and subsequent
inbound activity of the raw materials,
resources or other deliverables into the
manufacturing unit, warehouse, plants or
retail stores.
(ii) Outbound Logistics
In contrast, Outbound Logistics are
activities involved in storage and
subsequent outward movement of the
final product to the customers & end users
through the delivery channel.
3 (b). Nodes of Logistics Network
As described previously, a Logistics
Network resembles closely with a Neural
Network. They both contain nodes which
process the information entering it and
based on certain parameters provide
output or forward the information to the
next node. While a package / deliverable
travels from the start point ( supplier ) to
the end point ( customers or end users ), it
travels through various nodes where they
are either simply stored or further
processed and then forwarded again to
the next node. Each node has its own
prerequisites, capacity and cost involved.
9 | P a g e
9 Predicting Success of Logistics Network using Artificial Neural Network
In some cases there can be more than one
available routes for the transportation and
it becomes necessary for the business to
opt for most advantageous path for either
quick delivery or cost effective or both. The
nodes of a Logistics network can be:
a) Suppliers of goods or information
b) Factories or plants
c) Transit points
d) Warehouses & Storage Facilities
e) Distribution centers
f) Sales Agents / Brokers
g) Retail points
h) Disposal Centers
i) Customers or End Users
3 (c). Network Optimization
A network be it a Logistics Network or a
Neural Network is optimized to increase its
performance and reduce wastage of
resources whenever possible.
Optimization also leads to better decision
making and agile to fluctuations. Neurons
in our brain optimize themselves by
making new connections based on some
logic and train it for various scenarios.
Likewise a Logistics Network optimizes
itself by using various statistical tools
available to weigh the result in various
scenarios. The parameters of optimization
can be different in different cases such as
cost, distance, probability of fulfilling its
quota, etc.
3 (d). Network Optimization Algorithm
There are various algorithms and
methodologies available to optimize a
network. One of the most commonly used
algorithms is the Minimum Cost Rule
which takes into account the cost
associated in transporting from one
location to another and the path where
the cost is minimum is selected. In this
case, the path selected may be even
longer than some of the other paths or the
time taken by it larger. But, given the
company has sufficient time, it’s the most
cost effective algorithm.
One such other algorithm is the
Shortest Path Rule which takes into
account the shortest path between two
points or nodes. This algorithm is
preferred when the company is working
on tight deadline.
3 (e). Conclusion
It is evident from the from the literature
that to survive the competitive world of
business it is imperative to work on
Logistics network to make process more
cost effective and agile. However, the
Logistics can sometimes be simple costing
of only two nodes: starting node & finish
node. But sometimes it can be very
complex where network analysis becomes
very important and it is where Neural
Network and other algorithms come to
rescue. The parameters affecting the
network and the nodes can be bounded /
unbounded, simple / complex, single /
multi – dimension or small / large. And, as
the number of parameters increase, we
start to get the real picture. The size and
volume of realization becomes impossible
to work out using pen and paper.
Sometimes it becomes even hard to solve
it using traditional computing. It is where
MATLAB comes as a solution of all
problems. It is software which supports
Neural computing by default and easy to
program and create graphical
representation of analyzed data.
The detailed research analysis and
implementation is done in the following
sections working on a sample random
data generated by the software itself
based on a virtual Logistics Network. The
random data can also be replaced by
actual data it the necessity comes. The
paper goes on analyzing the network in
every aspect possible.
10 | P a g e
10 Predicting Success of Logistics Network using Artificial Neural Network
4. RESEARCH METHODOLOGY
It is known to everyone that the most
important aspect of any Supply Chain as
well as of Procurement process is the
“Delivery Network”, the network of nodes
through which transportation of goods &
materials take place. So by choosing a
network path where either the cost or the
distance is minimum half of the job is
already done. But it lacks the dynamic
character which exists in the real world
(such as fluctuations due to financial /
economic / political environment). So, in
the second part of the problem we work
with approximated dynamic setup with
several input parameters.
Data Source
A research is as good as its data. So, every
possible action was taken to keep the data
authentic without the presence of any
error. The data used in this project work is
randomly generated by the system, i.e.,
the data may be factually incorrect. It is
due to lack of that particular confidential
company data. Utmost care has been
given to keep the values of data within
bounds of the real world. But, these data
can be easily replaced by real world data
values keeping the algorithm and process
exactly the same. Since, the project work
only looks at the basic logistics setup and
only aims at providing working details
rather than specific analysis on a particular
company; so the data generated by the
system is sufficient to study and analyze
the neural network from educational point
of view.
The graphs and illustrations are
generated with the help of MATLAB
software which is best suited for use in
neural networks. The particular module
used is called Neural Network Tool. Only a
few available graph types are generated,
whereas people have the ability to
generate even more types of graphs and
various others cross combinations of data.
Each type of graph has their own
advantages and disadvantages.
Research Approach
Research is totally based on neural
network computing algorithms which are
an aspect of Management Information
System (MIS). MIS uses various categories
of analytical tools to take decisions in a
management situation. The project also
uses software coding algorithms to make
use of the neural network programming
already existing in the MATLAB software
suite. It also includes brain storming
sessions to generate algorithms for
shortest cost path and least distance path.
Testing & Evaluation
In the first part of the Research
Implementation raw data generated from
the shortest cost path is used. In the
second and final part of research, multi-
variable data is used. The accuracy of data
and its results is tested internally by the
software. However, the user has the ability
to determine the number of instances of
stimuli which he / she want to test in the
given neural network setting. The neural
network has three options namely: Train,
Validation and Test. In both validation
phase and test phase, inputs are given to
the neural network and test the
effectiveness of the output data.
So, se see that the project works
only with preliminary aspects of research
in this field. There are numerous scopes in
which the project can be improved. The
project work is an experimental piece of
work on testing effectiveness of neural
network in the field of management
analytics especially, the logistics network
and the supply chain network. In the
coming chapters we dive deep into
research implementation & sample
experimentation work.
11 | P a g e
11 Predicting Success of Logistics Network using Artificial Neural Network
5. RESEARCH IMPLEMENTATION
To start with, we start with a virtual
Logistics Network. The sample data is
randomly generated. Three matrices are
generated which consists of:
• Distance matrix
• Cost matrix
• Probability matrix
All the three matrices are 10 * 10 matrix
and the data is generated using MATLAB
random function which generates random
array of numbers. The distance matrix
represents the distance between two
nodal points and the cost matrix the cost
associated in transportation and storage
at nodal points. The probability matrix
takes care of the fluctuations and seasonal
variations which may appear in the cost
matrix. Each element of cost matrix is
added with each respective element of the
probability matrix to get the final cost
matrix. The function used to generate the
random data matrix is:
� = ������, �
Here, x and y represents the two
dimensions of the matrix. And, A is a x * y
matrix with all the individual values
between the range 0 and 1. To get the cost
and distance matrix, this function is
multiplied by a factor. For example:
� �������� = � ����������, � ∗ �����
Here, the Cost Matrix is multiplied by a
factor of 1000 and so all the values will lie
in the range greater than 0 but less than
1000. And, the round function converts the
value to the nearest one eliminating the
point places. Similar to cost matrix,
distance matrix is also generated.
���������� = � ����������, � ∗ ������
The values of cost matrix are given in INR
Lakhs while the values of distance matrix
are given in Kilometres. All the sample
matrices are given below. A : J and K : T
are the nodal points associated with the
Logistics Network.
A B C D E F G H I J
K 157 876 704 436 137 191 778 302 78 230
L 348 490 572 502 172 681 190 49 46 238
M 403 183 755 866 930 281 115 228 754 824
N 348 686 816 483 280 377 912 902 457 493
O 195 135 406 303 424 874 123 691 315 540
P 448 213 197 90 705 863 813 88 931 593
Q 49 585 545 623 970 426 130 608 536 827
R 449 735 499 349 306 382 123 309 394 465
S 394 929 484 347 166 373 117 217 946 122
T 920 50 578 226 856 921 281 642 449 114
Table 1. Cost Matrix of the Network ( in INR Lakhs )
A B C D E F G H I J
K 9902 8118 8714 4838 3715 7051 2264 4566 7323 7646
L 4050 3148 9050 8258 2579 2134 6258 7913 7862 2632
M 7632 5988 7207 6772 4614 8970 985 7752 6377 8345
N 3279 409 149 4312 1329 1167 1350 1649 1923 3983
O 6708 7000 5684 8071 447 8622 5591 3437 1225 4282
P 4008 9132 9210 978 2158 9716 59 987 4077 2300
Q 156 8180 8874 9058 7692 7317 1918 9327 2092 9023
R 5643 3793 9421 7838 3439 6654 8306 9531 7556 3602
S 6185 1514 8880 4199 6187 1339 7752 6814 9315 8787
T 1989 5090 883 7231 3419 8590 8974 9108 8724 501
Table 2. Distance Matrix of the Network ( in Kilometres )
12 | P a g e
12 Predicting Success of Logistics Network using Artificial Neural Network
A B C D E F G H I J
K 0.3577 0.4666 0.0629 0.0766 0.8989 0.9874 0.8403 0.2986 0.2891 0.8670
L 0.6813 0.9738 0.2661 0.3828 0.2271 0.8583 0.3547 0.1322 0.3943 0.5605
M 0.5432 0.0786 0.6342 0.7042 0.6898 0.2217 0.0438 0.0317 0.2321 0.6050
N 0.9957 0.6954 0.0035 0.8875 0.8866 0.8612 0.0489 0.4074 0.3844 0.8173
O 0.9386 0.1575 0.5219 0.9384 0.5309 0.1954 0.2673 0.9402 0.7235 0.2163
P 0.1314 0.1394 0.0279 0.4660 0.8255 0.0643 0.9378 0.6046 0.3836 0.6502
Q 0.9802 0.9147 0.5725 0.9511 0.2246 0.7436 0.7786 0.1367 0.3118 0.7136
R 0.6398 0.1863 0.5436 0.8118 0.2646 0.0592 0.4683 0.8497 0.2046 0.2442
S 0.4832 0.8941 0.4781 0.7825 0.7164 0.4231 0.1120 0.7354 0.2084 0.5437
T 0.7584 0.4073 0.6258 0.7982 0.8239 0.3393 0.5348 0.9655 0.7821 0.7647
Table 3. Probability / Noise Matrix
������������� ( , �) = ������� ( , �) + �����(��������������� ( , �) ∗ ���)
A B C D E F G H I J
K 193 923 710 444 227 290 862 332 107 317
L 416 587 599 540 195 767 225 61 85 294
M 457 191 818 936 999 303 119 231 777 885
N 448 756 816 572 369 463 917 943 495 575
O 289 151 458 397 477 894 150 785 387 562
P 461 227 200 137 788 869 907 148 969 658
Q 147 676 602 718 992 500 208 622 567 898
R 513 754 553 430 332 388 170 394 414 489
S 442 1018 532 425 238 415 128 291 967 176
T 996 91 641 306 938 955 334 739 527 190
Table 4. Actual Cost Matrix ( in INR Lakhs )
We take the above example and try to
implement the shortest path algorithm to
the cost matrix as well as the distance
matrix. By applying the algorithm to cost
matrix we get the least coast involved in
the transportation and the route through
which it needs to be travelled while in case
of distance matrix the total cost of
transportation along the route with least
distance is calculated.
Suppose that a material needs to
travel from a node point to another node
point ( or suppose from a supplier to
manufacturer / customer via various
storage location or transit points ). And,
there is a need to find the following
routes:
a. Through which cost is minimum
b. Through which distance travelled is
minimum
For that we need to feed the above matrix
into the program and we get the result as
output. Various operators and functions
used for preparing the coding of this
algorithm are as follows:
• rand( x , y ) : generates a random
matrix of x , y dimension
• randi([ x , y ]) : generates a random
integer between the range of x , y
• clear all : clears all the local
and temporary variables in memory
• clc : clears the command
window
• size( x ) : shows the size or
dimension of a matrix x
• for / while : loop statement
• if / else if : condition statement
• end : ending sequence for
loop statement or conditional
statement
13 | P a g e
13 Predicting Success of Logistics Network using Artificial Neural Network
The complete MATLAB code for finding the
shortest path is given below.
clear all; clc;
XM=10;YM=10; X=1;Y=1; a=round(rand(XM,YM)*10); results=[];
length=0;
for z=1:10000 x=X;y=Y; temp=a(x,y); r1=[1];r2=[1]; rtemp=[a(x,y)];
while (x<XM | y<YM)
xn=x;yn=y; while (xn==x & yn==y) if (x<XM & y<YM) xr=[x x+1];yr=[y y+1];
elseif (x==XM & y<YM) xr=[x];yr=[y y+1]; elseif (x<XM & y==YM) xr=[x x+1];yr=[y]; end
x1=size(xr); y1=size(yr);
xn=xr(1,randi([1,x1(1,2)])); yn=yr(1,randi([1,y1(1,2)])); end
x=xn;y=yn;
temp=temp+a(x,y); rtemp=[rtemp,a(x,y)]; r1=[r1,x];r2=[r2,y]; end
if (temp<length | length==0) results=rtemp; length=temp;
end
end a
length results
In the code, a is a randomly generated
matrix. But for practical purposes we can
feed the code by replacing it with actual
data. XM & YM are the dimensions of
matrix. So, also we need to change these
with the change in number of nodes. At
the bottom; a, length and results is used
to provide the output of the algorithm
where a represents the actual matrix,
length gives the total cost involved ( in
case of cost matrix ) or total distance ( in
case of distance matrix ) and results gives
the output path along the least cost /
distance.
Now let us feed the actual cost
matrix in the system and get the output.
a =
157 876 704 436 137 191 778 302 78 230
348 490 572 502 172 681 190 49 46 238
403 183 755 866 930 281 115 228 754 824
348 686 816 483 280 377 912 902 457 493
195 135 406 303 424 874 123 691 315 540
448 213 197 90 705 863 813 88 931 593
49 585 545 623 970 426 130 608 536 827
449 735 499 349 306 382 123 309 394 465
394 929 484 347 166 373 117 217 946 122
920 50 578 226 856 921 281 642 449 114
length = 3841
results = [ 157 348 183 816 303 705
426 123 217 449 114 ]
The final path would be:
157 876 704 436 137 191 778 302 78 230
348 490 572 502 172 681 190 49 46 238
403 183 755 866 930 281 115 228 754 824
348 686 816 483 280 377 912 902 457 493
195 135 406 303 424 874 123 691 315 540
448 213 197 90 705 863 813 88 931 593
49 585 545 623 970 426 130 608 536 827
449 735 499 349 306 382 123 309 394 465
394 929 484 347 166 373 117 217 946 122
920 50 578 226 856 921 281 642 449 114
So, it is evident from the shortest path
algorithm when applied to the cost matrix
is that the minimum project cost for
transportation would be INR 3841 lakhs.
The work is only half done. The next step
would be to analyze shortest path
algorithm on distance path and to see
whether both the paths coincide for cost
and distance matrix.
14 | P a g e
14 Predicting Success of Logistics Network using Artificial Neural Network
Distance matrix, a =
99
02
81
18
87
14
48
38
37
15
70
51
22
64
45
66
73
23
76
46
40
50
31
48
90
50
82
58
25
79
21
34
62
58
79
13
78
62
26
32
76
32
59
88
72
07
67
72
46
14
89
70
98
5
77
52
63
77
83
45
32
79
40
9
14
9
43
12
13
29
11
67
13
50
16
49
19
23
39
83
67
08
70
00
56
84
80
71
44
7
86
22
55
91
34
37
12
25
42
82
40
08
91
32
92
10
97
8
21
58
97
16
59 98
7
40
77
23
00
15
6
81
80
88
74
90
58
76
92
73
17
19
18
93
27
20
92
90
23
56
43
37
93
94
21
78
38
34
39
66
54
83
06
95
31
75
56
36
02
61
85
15
14
88
80
41
99
61
87
13
39
77
52
68
14
93
15
87
87
19
89
50
90
88
3
72
31
34
19
85
90
89
74
91
08
87
24
50
1
length = 47555
results = [ 9902 3148 5988 149
4312 1329 1167 5591 987
2092 3602 8787 501 ]
The final path would be:
99
02
81
18
87
14
48
38
37
15
70
51
22
64
45
66
73
23
76
46
40
50
31
48
90
50
82
58
25
79
21
34
62
58
79
13
78
62
26
32
76
32
59
88
72
07
67
72
46
14
89
70
98
5
77
52
63
77
83
45
32
79
40
9
14
9
43
12
13
29
11
67
13
50
16
49
19
23
39
83
67
08
70
00
56
84
80
71
44
7
86
22
55
91
34
37
12
25
42
82
40
08
91
32
92
10
97
8
21
58
97
16
59 98
7
40
77
23
00
15
6
81
80
88
74
90
58
76
92
73
17
19
18
93
27
20
92
90
23
56
43
37
93
94
21
78
38
34
39
66
54
83
06
95
31
75
56
36
02
61
85
15
14
88
80
41
99
61
87
13
39
77
52
68
14
93
15
87
87
19
89
50
90
88
3
72
31
34
19
85
90
89
74
91
08
87
24
50
1
The yellow colour shows the path
traversed by the network while green
colour shows the starting and the ending
node. The shortest distance between the
nodes is 47555 kms. Hence, from the two
marked matrices it can be noticed that the
shortest paths are different in both the
cases. So, network optimization is required
to make both the path same. From
Logistics point of view, it’s best for
business that both the paths are same. If
both the paths are same then materials
take minimum time travelling it while the
cost involved is also minimum. The above
example can be viewed as a Logistics
network of Rail / Road / Air transit points
where flow from one location to another is
associated with two parameters namely
cost and distance. And, we worked to find
the best route available for the
transportation of goods / information.
Now let us take some time to give a brief
overview of the MATLAB coding hereby
used.
The code calculates the shortest
path between the top left corner and the
bottom right corner. At first marker is set
at the top left corner. At that point it has
three places to go, i.e. place directly below
it, the place directly right to it or the place
which is at diagonal. The choice of this is
done randomly. Even while a marker has 8
places surrounding it, then also 3 places
are considered which is towards the
bottom right corner. As the marker
reaches the bottom right corner, one cycle
is completed. Since this code works on
random position on cursor hence the
entire procedure is repeated for a large
number of cycles. Whenever a cycle has a
lower value it is replaced by previously
held value. Hence, at the end of complete
running of the code, the shortest distance
is found with more than 99% accuracy.
This program works on matrix of any
dimension but a higher dimension matrix
would take more time and more
computing resources to complete
successfully.
The shortest path algorithm is like a
Neural Network where the data takes
random path from source to destination.
The best solution in this process is
recorded. These types of neural networks
15 | P a g e
15 Predicting Success of Logistics Network using Artificial Neural Network
are known as Stochastic Neural Networks.
They work by introducing random
variation in the network which makes
them perfect to be used for network
optimization. Hence, in the example
above; simulation of the network was
done to find the shortest cost path and the
shortest distance path.
After we get the individual result of
each matrix, now the next step is to collect
all the input output pair in a different
matrix. Now we take the cost matrix and
work further on it. We take the
determinant value of each cost matrix and
insert it into a different input vector; and
the corresponding length in another
output vector. We gather input output pair
for 1000 random cost matrices. This input
output pair is then fed into the MATLAB
neural network simulator. To generate the
input output pair, the following coding is
used:
clear all;
clc;
testInput=[];
testTarget=[];
XM=10;YM=10;
X=1;Y=1;
for p=1:1000
a=round(rand(XM,YM)*1000);
b=round(rand(XM,YM)*10000);
results=[];
length=0;lx=0;
for z=1:100
x=X;y=Y;
temp=a(x,y);l2=b(x,y);
r1=[1];r2=[1];
rtemp=[a(x,y)];
while (x<XM | y<YM)
xn=x;yn=y;
while (xn==x & yn==y)
if (x<XM & y<YM)
xr=[x x+1];yr=[y y+1];
elseif (x==XM & y<YM)
xr=[x];yr=[y y+1];
elseif (x<XM & y==YM)
xr=[x x+1];yr=[y];
end
x1=size(xr);
y1=size(yr);
xn=xr(1,randi([1,x1(1,2)]));
yn=yr(1,randi([1,y1(1,2)]));
end
x=xn;y=yn;
temp=temp+a(x,y);
rtemp=[rtemp,a(x,y)];
r1=[r1,x];r2=[r2,y];
end
if (temp<length | length==0)
results=rtemp;
length=temp;
lx=l2;
end
end
testInput=[testInput,lx];
testTarget=[testTarget,length];
end
testInput
testTarget
The following code works with a cost
matrix of XM , YM dimension and gives the
output matrices of testInput & testTarget.
This output is then fed into MATLAB Neural
Network Simulator. The simulator is started
by typing in the command nnstart in the
command window, which starts a
simulator with a GUI interface.
After starting the GUI interface we
follow these steps to perform further
optimization and Neural Network Setup.
a) Select Dynamic Time Series App
b) Select Nonlinear Autoregressive
with External Input (NARX)
c) Click Next
d) In Input dropdown select testInput
variable
e) In Targets dropdown select
testTarget variable
f) In Time step select Matrix Column
g) Click Next
16 | P a g e
16 Predicting Success of Logistics Network using Artificial Neural Network
h) Select Validation and Testing
percentage to 20 % each. The
Training will become 60 % by itself.
i) Click Next
j) Change Number of Hidden Neurons
to 30 and number of delays to 5.
k) Click Next
l) Now choose a training algorithm.
We chose Levenberg-Marquardt.
m) Click on Train
n) Generate the Performance Graph
o) Generate Error Plot, Correlation
Graph & other plots whichever is
necessary.
p) Click on Next
q) Under Optionally Perform
Additional Tests again select the
testInput matrix for Input field and
testTarget for Target field.
r) Click Test Network
s) Click on Next
t) Click on Neural Network Diagram
u) Finally click on Generate MATLAB
Function to generate the final code
for the Neural Network.
So, the testInput matrix and the
testTarget matrix which contains the input
and target pair has the dimension of 1 * n.
Both the matrices contain data in the
following format:
testInputtestInputtestInputtestInput = 1.0e+291.0e+291.0e+291.0e+29 **** [[[[----0.3312 0.3312 0.3312 0.3312 ----0.8003 0.8003 0.8003 0.8003
----0.1425 2.9128 0.1425 2.9128 0.1425 2.9128 0.1425 2.9128 ----0.1145 0.1145 0.1145 0.1145 ----0.8028 0.8028 0.8028 0.8028
----0.6527 0.0303 0.6527 0.0303 0.6527 0.0303 0.6527 0.0303 ----0.0609 0.1297 0.0609 0.1297 0.0609 0.1297 0.0609 0.1297
----0.1761 0.0371 0.1761 0.0371 0.1761 0.0371 0.1761 0.0371 ... ... ... ... ... ... ... ... ............ ]]]]
testTargetestTargetestTargetestTarget = [4375 5581 3289 = [4375 5581 3289 = [4375 5581 3289 = [4375 5581 3289
5817581758175817 4073 5860 5046 5936 4073 5860 5046 5936 4073 5860 5046 5936 4073 5860 5046 5936
5698 6386 5698 6386 5698 6386 5698 6386 ............ ... ... ... ... ...]...]...]...]
One of the main characteristics of a Neural
Network is adaptive learning. It learns
during the training period. So, if the
network is trained for multiple times, its
efficiency slowly increases. The code for
Neural Network hence generated can be
used as a function in some other
programs without the need of repeating
the tasks described again and again. The
code generated in given below:
function [Y,Xf,Af] =
myNeuralNetworkFunction(X,Xi,~)
%MYNEURALNETWORKFUNCTION neural network
simulation function.
% Input 1
x1_step1_xoffset = -4.98999253881676e+29;
x1_step1_gain = 2.43320745266833e-30;
x1_step1_ymin = -1;
% Input 2
x2_step1_xoffset = 2389;
x2_step1_gain = 0.000415627597672485;
x2_step1_ymin = -1;
% Layer 1
b1 = [1.8622064538453846;-
1.3174394252249615;1.9871803820807037;2.00
36257151714532;1.5905552659314099;-
1.7757453601101036;-1.133700114441897;-
1.1610975421466321;1.0547846028904724;-
1.209130730342453;0.21669830424588038;0.54
201616038863742;0.9063674742950677;-
0.087828013925516285;-
0.33129153301616143;-
0.3102409292987372;0.36928862994813999;0.4
3016239621204977;0.17090046832086914;0.728
28274161038487;0.49848283419706602;-
0.48909627613539775;0.98650826861617913;1.
0342303883301023;0.91787242299088112;-
1.5080617045328351;-1.076145305423639;-
2.3354188909909031;2.4260369891928395;1.63
56262790583684];
IW1_1 = [-0.8990662879531256
0.63461441145057462 0.62069385836740265
0.73868345515978784
0.18065175069981662;0.83514571633870671
0.49432518088486799 0.16614340202975794
0.64806118134276736 0.12905628337033456;-
0.026496114337601035 0.13954316354780882 -
1.1845489186968488 0.58886316300453012
0.46775714406894559;0.24337221641342624
0.27993231322857781 -0.61434779168579279 -
1.2497150186046722 0.63585373554834046;-
0.013105345224936202 0.38984235186004551 -
0.15806467311727126 -1.1028179320429057
0.21237483678855992;0.23814126292331
0.5119158762223095 0.078313953593802083 -
0.96839287378850569
0.261817165927826;0.38435356615316235
0.15795276738597055 -0.99790964039322505
0.65885836132524589 0.61652792897625142;
0.39757621067164434 -0.486336538715748 -
0.19126431183574655 -0.91834363610115655 -
0.75615096639269364;-1.3748683911122817
0.10833156118945539 -0.37291015867472049
0.084893388233545947 -
0.027105282183725349;1.1455673347416533
0.53003726632197057 -0.47084124912411263
0.54119986676313814 -0.52968706330652982;
0.95276031958611007 0.29468288623147559 -
0.0063621284686704999 -2.6488401124019521
17 | P a g e
17 Predicting Success of Logistics Network using Artificial Neural Network
0.81553830857503695;-0.21869342802252509
0.23359834929142709 0.55908184351159917
0.20247696276611285 -0.89166792247530846;-
0.19219735592111095 -0.6929442795969627
0.44413234208284058 0.028171323570738566
0.52867998445926723;-0.80544929429151213 -
0.70960043978053555 -0.25668434724595557 -
0.74040111346637461
0.65513673328972544;0.72257785097141947
0.59687143961080269 0.67487290163600855 -
1.0503050331803363
0.58705665912220306;0.34309769040657734
0.57908827555917342 -0.92254795366028486 -
0.81037216649739419 -
0.19635804242575949;0.9384730365578271
0.079036443022610689 -0.033884509095074913
0.0086077331698611037 -
0.54936897550401143;0.33914637239829887 -
0.38940273136201964 0.23933785195616014
0.75203180951076187 -
0.26275431507056585;0.5033458562545432
0.20986094044036369 2.6370533627181731 -
0.93046359659903666
0.85836339352292179;0.70745880175925957 -
0.14445441321995967 -0.44112025728102122 -
0.98091735881663544 -1.2632184679735374;-
0.28762577820823498 -1.0142405287027487 -
0.46376470886144222 -0.31176220871814775
1.3246741992758304;0.52418104732678672
0.2762184661076818 1.4821886659486661 -
1.0466682714588009 -
0.41634183406386549;0.97039527504630074
0.4704524071147253 0.74280607436066359
0.69702490138182704 0.23983591620954248;-
0.55912436050875702 0.59830234669057969
0.65300134116136177 0.44186425739382568 -
1.0795960947346597;0.47835467047887176 -
0.88369549718136475 -0.2078912740824943 -
0.74222955412495439 -0.56068541374555247;-
0.64203716571715497 0.32288247956151234 -
0.30317220206179352 -0.98654247000060336
0.80927029855826194;-0.36624915607711345 -
0.77702853510993908 0.72215093887452553
0.71303583113890234 0.66120508707766668;-
0.53024896849301084 -0.17212434682316929 -
0.00044656499703289373 0.54849998163047564
0.70620579740084444;0.47978752559878707
0.63456884801241331 0.14239686178803643
0.81578454678473844
0.91246619454709377;0.84229378752886486 -
0.54899708465557717 0.99474871880497606
0.79640279449997364 -
0.037553398680422921];
IW1_2 = [-0.34611344785693254 -
0.40830675831848545 -0.86085433308025094
0.96480333925503114
0.37266981117479286;0.73536093575949479 -
0.77041359053094483 -1.0621340035067459
0.091383347730121106
0.8020243215829721;0.51949901018363076 -
0.08583137819447699 0.39445447714986326
0.94575351747801806 -0.0917014538018429;-
0.56843871876638952 -0.42478948686014029
0.17908729598308123 -0.1813913782930516
0.52028997376900543;0.24173913466964367 -
0.019117405919199054 0.79082190259439999 -
0.53989605123072049 0.68566975800058361;-
0.82493960622365103 0.77488597809783832 -
0.09972668064817361 -0.22626322996665793
1.2235094452239581;1.3325546981674625
0.04084751033621372 0.81272081662096896 -
0.68882196442835197 0.42526157859027569;-
1.1624456301620405 0.074752836632442365
0.28698793803347228 -0.18836171731554191
1.6994328309264659;0.082641165505694864 -
0.72647638294345307 0.37222916618193874 -
0.59557182059838309 -
0.54264118200755551;0.6170121320454558
0.51443721915202689 -0.13301985987703244
0.63812925800480447 -
0.017084794628528632;0.0017130577130541197
-0.72800368520101699 0.10703330995541277 -
0.17316408790626062 -
0.3755092278573019;0.35223039924279559 -
0.13764491172306542 -1.13361136467351
1.5277529899169502
1.7189480228176701;1.1672419360255182 -
0.10334655171986172 0.53427977367682888
1.5139044180624517 -1.550737677175634;-
1.0569191856627287 -0.26358498521492124 -
1.0464774105351735 0.59866436090761566
0.16687822195650651;-1.2789849892807565 -
0.12786975087607283 -0.02558122556063841
0.80215017670022715
0.55039043808640442;0.61421688780853689 -
1.0028508140191952 0.17348480707386513
0.87214976193028992 1.0579861500912122;-
1.5529209822890586 -0.43808715501915063 -
0.59074272848153941 0.88027658657167618
0.86266064323177183;-0.05043044769732264
0.47044464482943216 -0.12667436822643174
0.37517910056700515
1.043301418892971;0.47022588543937671 -
0.65691194612863257 -0.97086448280318516
0.094226358566928708 -
1.1293816066309033;1.0421132922519136
0.065357225366302646 -0.94145327173758075
0.86012002196354642 0.28503649126835584;-
0.48094833405868009 -0.82269501151454727
1.1468223251793694 0.60030284193955996
0.17448178916311574;0.013868739336124189
0.98211319217429494 -1.4017692964368589 -
0.36548314935748533 0.39642465685040162;-
0.53132671320011282 -1.0596004121598877 -
0.095853903277147712 -0.30715393072113406
0.38775546556609308;0.29364552417358758 -
0.57991963812864633 0.1391714880152958
0.30755397521125971
0.1327141859513469;0.4226223163090283 -
0.56216412131062587 0.9658677372049681 -
0.90499507143959779 0.2589082153187397;-
0.53149722899416429 -0.69210825201258275 -
1.4810029456546885 -0.81809206139490032
1.2776826328475663;-0.12991100698731972
0.77426373433953333 -0.44820799882802415 -
0.92006271632766257 -
0.37847869211162261;0.80518593508973324 -
0.48792315659912644 0.99758979395753478
18 | P a g e
18 Predicting Success of Logistics Network using Artificial Neural Network
1.4785751355499472 -0.035433985779074202;-
0.46555554288242873 -0.057754331726513439
0.11789196263562859 -0.5617207144591243
0.040644474042050689;-0.59831406066985027
-1.1672947553207704 -0.3915596324049469
0.84081490926390678 0.2556588847771385];
% Layer 2
b2 = 0.08711061819165343;
LW2_1 = [0.45852110202363688
0.12637545383546009 -0.30321962693291749 -
0.46561838228245878 -0.60165853483385245 -
0.6141655036760979 -0.33844568110575796 -
0.89559170822650036 0.51470320408859283
0.12788968205527926 -0.43402497270887641
0.17947011010088426 -0.55226708854813755 -
0.12813696032788227 0.80011444752262162
0.20648492637305527 -0.77525379331602473 -
0.41751528644098324 -0.36547039909167822
0.18857855743724139 0.45470559546583444
0.48652515383598738 0.22978250493786384 -
0.70688834235757514 0.36382965901896747 -
0.45786941916913054 -0.06523296931512948 -
0.5769283845107902 -1.0834250534825229
0.42348378732602371];
% Output 1
y1_step1_ymin = -1;
y1_step1_gain = 0.000415627597672485;
y1_step1_xoffset = 2389;
% ===== SIMULATION ========
% Format Input Arguments
isCellX = iscell(X);
if ~isCellX, X = {X}; end;
if (nargin < 2), error('Initial input
states Xi argument needed.'); end
% Dimensions
TS = size(X,2); % timesteps
if ~isempty(X)
Q = size(X{1},2); % samples/series
elseif ~isempty(Xi)
Q = size(Xi{1},2);
else
Q = 0;
end
% Input 1 Delay States
Xd1 = cell(1,6);
for ts=1:5
Xd1{ts} =
mapminmax_apply(Xi{1,ts},x1_step1_gain,x1_
step1_xoffset,x1_step1_ymin);
end
% Input 2 Delay States
Xd2 = cell(1,6);
for ts=1:5
Xd2{ts} =
mapminmax_apply(Xi{2,ts},x2_step1_gain,x2_
step1_xoffset,x2_step1_ymin);
end
% Allocate Outputs
Y = cell(1,TS);
% Time loop
for ts=1:TS
% Rotating delay state position
xdts = mod(ts+4,6)+1;
% Input 1
Xd1{xdts} =
mapminmax_apply(X{1,ts},x1_step1_gain,
x1_step1_xoffset,x1_step1_ymin);
% Input 2
Xd2{xdts} =
mapminmax_apply(X{2,ts},x2_step1_gain,x2_s
tep1_xoffset,x2_step1_ymin);
% Layer 1
tapdelay1 = cat(1,Xd1{mod(xdts-[1 2 3
4 5]-1,6)+1});
tapdelay2 = cat(1,Xd2{mod(xdts-[1 2 3
4 5]-1,6)+1});
a1 = tansig_apply(repmat(b1,1,Q) +
IW1_1*tapdelay1 + IW1_2*tapdelay2);
% Layer 2
a2 = repmat(b2,1,Q) + LW2_1*a1;
% Output 1
Y{1,ts} =
mapminmax_reverse(a2,y1_step1_gain,y1_step
1_xoffset,y1_step1_ymin);
end
% Final Delay States
finalxts = TS+(1: 5);
xits = finalxts(finalxts<=5);
xts = finalxts(finalxts>5)-5;
Xf = [Xi(:,xits) X(:,xts)];
Af = cell(2,0);
% Format Output Arguments
if ~isCellX, Y = cell2mat(Y); end
end
% Map Minimum and Maximum Input Processing
Function
function y =
mapminmax_apply(x,settings_gain,settings_x
offset,settings_ymin)
y = bsxfun(@minus,x,settings_xoffset);
y = bsxfun(@times,y,settings_gain);
y = bsxfun(@plus,y,settings_ymin);
end
% Sigmoid Symmetric Transfer Function
function a = tansig_apply(n)
a = 2 ./ (1 + exp(-2*n)) - 1;
end
% Map Minimum and Maximum Output Reverse-
Processing Function
function x =
mapminmax_reverse(y,settings_gain,settings
_xoffset,settings_ymin)
x = bsxfun(@minus,y,settings_ymin);
x = bsxfun(@rdivide,x,settings_gain);
x = bsxfun(@plus,x,settings_xoffset);
end
19 | P a g e
19 Predicting Success of Logistics Network using Artificial Neural Network
The efficiency of a Neural Network can be
observed from the MSE ( Mean Squared
Error ) value of the network. The lesser the
MSE value, the better is the performance.
To analyze the flow of MSE value as the
system is simulated, we need to look at
the MSE graph which is given below.
Fig 6. MSE graph for 1
st Training period
From the graph, it is noticed that Best
Validation Performance is at epoch 7
which means that the network reached its
peak performance on the 7th run cycle.
But, further cycles were done to test and
validate the data. If the network is trained
again for the 2nd time, we get the following
graph.
Fig 7. MSE graph for 2
nd Training period
From the graph it is noticed that in the 2nd
training period, peak performance is
obtained in the epoch 4 which is 3 cycles
better than the last training. Hence, the
common observation would be that the
performance of the neural network
increase with each training cycle. Also
notice that the peak MSE value is also
reduced. All these prove that the network
adapted to the incoming stimuli.
Fig 8. Training State of ANN
Fig 9. NARX Neural Network Diagram
The neural network as compiled in the
steps given above is shown in Fig 9. It
works on single input system x(t) and gives
the output y(t) which is again fed into the
network. The number of hidden layers is
30 and the amount of delay is 5. The
output layer is the activation layer which
decides if there will be output based on
activation function, input vector and
weights. The total sample is first divided
into three parts: training data, test data
and validation data ( here 60% of the
sample are used for training, 20% for
testing and remaining 20% for validation ).
Training is first done to educate the
network to incoming stimuli. Then testing
is done to test the performance of the
network. Finally, validation is used to make
sure that the network does not show any
abnormal behaviour and that it is working.
20 | P a g e
20 Predicting Success of Logistics Network using Artificial Neural Network
Let us now analyze the MSE graph. The
validation curve closely matches to the test
curve which means there is no over fitting.
Contrary, if the test curve had been ahead
of the validation curve, then some over
fitting might have occurred.
Fig 10. Regression Plot for the Network
The graph shows the relationship between
the shortest cost path in the cost matrix
and the respective distance from the
distance matrix. Hence, it shows the
variability of cost associated with
transporting something between two
nodes of certain distance. Since, the data
was randomly generated without any
actual relevance to the world; hence
association between the two vectors is
minimum, almost equal to zero. R
represents the relationship between the
inputs and the targets. The dashed line
represents the perfect target which is
results – targets. The solid coloured line
represents the best fit. In this example, R =
0.19444 ( for combination of Training, Test &
Validation data ).
�� = �. �������
Hence, R2 = 3.78% .
So, there is least
amount of connection
between the two
variables. The points
which are scattered
across the graph
shows the data points
used as inputs for
training, validation
and test purposes. If
for a particular run
cycle, the network
does not yield good
results, then it is
recommended to run
the whole network
again as each cycle of
initialization involves
different network
parameters and so it
might produce
different results. We
can also change the
amount of hidden networks involved in
the network as each extra hidden network
gives the whole network to optimize itself.
We can also change the algorithm to see if
the yields are successful.
In the above text we worked briefly
with a simple example as to how the
neural network is setup and analyzed.
Hence the following can be inferred from
the data graphs given above:
• Data was based on a declining /
faulty / dead company
• Algorithm used was incorrect
• Data was bogus
• Data was manipulated for some
reason
21 | P a g e
21 Predicting Success of Logistics Network using Artificial Neural Network
The above example of Neural Network
was done using nnstart tool which has
scope for generation of the network
and other related analytics. It is also
used to generate the function which
can be later used to make predictions.
Now we would work with another
Neural Network tool nntool which can
be used effectively to make predictions
also. In the previous example data was
randomly generated hence there was
little amount of correct / appropriate
relationship in it. We again try the
whole neural network with relevant
data.
For the second example we take a
multi parameter matrix for the logistics
network.
Sl. Parameter Value Range
1 Noise 0.0063 – 0.99
2 Seasonal 0 / some value
3 Excise Rate Less than 10
4 Normal / Tatkal 0 / 1
5 Failure Rate 0.2 – 0.6
6 Inflation Rate Less than 10
7 % transported <=100
8 Distance (10000
miles)
<=5
9 No. Of nodes <10
10 Quantity <500
11 Approx Cost
(INR Lakhs)
<50
12 Time (days) <500
13 Overhead Cost
(INR Thousands)
<30
14 Output Cost ---------------------
The test matrix contains a single value pair
based on which prediction will be made.
Finally the predicted value is multiplied by
the multiplying factor for the cost matrix.
Only the data pair in the form of a matrix
is fed into the network. The neural
network automatically analyzes the data
column and trains the network which is
then reinforced automatically for the
situation by the AI. The predicted value will
be compared with company financial data
to get the final conclusion if the logistics
network is a case of success or fail. The
data pair used here contains more than
500 instances. Each instance contains 13
input parameter and 1 output parameter.
The input matrix is stored in testInput
(13*506 matrix) and the output matrix in
testTarget (1*506 matrix). The steps
followed for this example are given below:
a) We start the Neural Network / Data
Manager by typing in nntool in the
command window.
b) Click on Import.
c) Select testInput variable, select
destination as Input Data and click
on Import. Finally Click Ok.
d) Select testTarget variable, select
destination as Target Data and click
on Import. Finally Click Ok.
e) Select testData variable, select
destination as Input Data and click
on Import. Finally Click Ok.
f) Click Ok.
g) Click on Close.
h) In the Neural Network / Data
Manager Click on New
i) Select Network type. We select it as
Feed-forward backprop as the
network algorithm.
j) Select testInput as Input Data.
k) Select testTarget as Target Data.
l) Click on Create.
m) Click on Ok and then on Close.
n) Select network1 in the Neural
Network Manager and click on Open.
o) Select Train Tab. Under it select
testInput as Inputs & testTarget as
Targets. Finally Click on Train
Network.
p) Generate the Regression and
Performance Plots.
22 | P a g e
22 Predicting Success of Logistics Network using Artificial Neural Network
q) Click on Close and select the
Simulate tab.
r) Select testData as Inputs and click
on Simulate Network. Finally click
on Ok.
s) Under Neural Network Manager
select network1_outputs and click
on Open.
t) The predicted values are displayed.
Fig 11. Performance Graph for 2
nd example
Fig 12. Gradient, mu and validation fail plot
for 2nd example
From the above two graphs it can be
observed that MSE value at best
performance is 15.0921 which is quite
small. Comparing the MSE value for the 1st
example which was 621031 for 2nd training
cycle, we can clearly see the difference. We
can therefore conclude that the data has
all the merits. And, the best performance
was observed at epoch 10. From the
gradient curve it is observed that the error
rate was slowly reduced as the network
training was simulated. And, validation
checks were performed from epoch 10
onwards as the best performance was
observed in epoch 10. Now let us look at
the regression curve showing the data
points and fit curve.
Some of the sample values used to train
the network is given below:
1111 0.111 0.173 0.279 0.179 0.113 0.435
2222 0 0 0 0 30.00 0
3333 27.740 9.690 9.690 9.690 4.930 10.59
4444 0 0 0 0 0 1.00
5555 0.6090 0.585 0.585 0.585 0.428 0.489
6666 5.983 5.707 5.926 5.670 6.897 5.344
7777 83.50 54.00 42.60 28.80 54.30 100
8888 2.109 2.381 2.381 2.798 6.336 3.87
9999 4.000 6.000 6.000 6.000 6.000 4.00
10101010 711.0 391.0 391.0 391.0 300 277.0
11111111 20.10 19.20 19.20 19.20 16.60 18.6
12121212 396.9 396.9 396.9 393.2 391.2 396.
13131313 13.350 13.590 13.590 17.60 11.38 23.0
14141414 23.100 19.700 18.300 21.200 22.00 20.00
And the testData, the matrix which is used
to predict the scenario is also given
Sl. No. Values 1 0.2300
2 10.0000
3 10.5000
4 1.0000
5 0.5000
6 5.7000
7 68.0000
8 2.3500
9 6.0000
10 650.0000
11 20.0000
12 15.0000
13 22.0000
On running the simulation of the neural
network representing the logistics
network, the above set of data gave an
output of 1*1 matrix representing the cost
associated with the Logistics network.
����� = ��. ����
23 | P a g e
23 Predicting Success of Logistics Network using Artificial Neural Network
From the parameter table, the value of
output can be taken as INR 25.5999 Lakh.
We now look at the regression plot and
analyze it with respect to the regression
plot obtained in the first example.
Fig 13. Regression plot for 2nd example
From the above regression plot, the R-
value can easily be noted to be 0.92248
which is clearly better than the R value of
the previous example. Hence, the data
points are better fit. Even then some of the
data points are scattered around the fit
curve. Since, the R value is greater than 0.9
hence the network can be said to be
similar to that of the Logistics network and
that the data points are accurate.
The output value that we received
as a result of the prediction is the total
cost involved with the network for the
transportation under the parameters
given as the input. So, to decide if the
network is a case of success or fail, we
must compare the output cost ( INR
25.5999 lakhs ) to the amount of money
the company is capable and willing to
spend. If the output is less than that sum
of money, then the
network is a successful
one or else it is a case of
unsuccessful network.
The MATLAB code for
the neural network
which was generated in
this example is given
below.
function [Y,Xf,Af] =
neural_function(X,~,~) %NEURAL_FUNCTION
% ===== NEURAL NETWORK
CONSTANTS =====
% Input 1 x1_step1_xoffset =
[0.00632;0;0.46;0;0.385;
3.561;2.9;1.1296;1;187;
12.6;0.32;1.73];
x1_step1_gain =
[0.022479517787368;0.02;
0.0733137829912024;
2;4.11522633744856;
0.383215175320943;
0.0205973223480947;
0.181869435931944;
0.0869565217391304;0.00381679389312977;0.2
12765957446809;0.00504311866458218;0.05518
76379690949]; x1_step1_ymin = -1;
% Layer 1 b1 =
[36.357184522980909;14.119396930672583;-
0.27278614037853477;3.4046596216055507;-
11.269335472679851;-13.169119619915456;-
0.88236889557629072;23.938045111726254;0.3
2849481229537225;12.090640960323316]; IW1_1 = [-6.4156964643184944
12.916636442378659 36.160509774436825 -
14.47529888885345 20.203037621955211 -
17.681922329162976 21.070309153982532
27.994332023180903 23.1837743628844
1.7416235328037739 17.293024630348778 -
20.695869400746105 -
12.413148760083043;19.887374545364036 -
40.817613391351415 -3.6829731767938467 -
17.602203115150555 -117.7853249654877 -
94.008539298530934 27.923670320088508
44.50977196623667 57.395440325422641 -
40.172046845310021 18.314427951551593 -
74.156876361418739 -
22.038089268098158;22.845485688330754 -
24 | P a g e
24 Predicting Success of Logistics Network using Artificial Neural Network
7.0115781538036401 -0.0091110059063216758
-22.108140682162873 -1.3984949157611064
13.877773903176005 -4.8939641488803893
14.65222686408784 -8.0655091321759276
2.7385798751957333 -1.1196236353942755
9.6334073350981946
20.41669141826533;1.072620238762759 -
0.30693381329168934 0.46051583907811083
0.28340723772764953 0.51158114965318713
0.20688380402814513 -0.23617158155962514
1.9523863679811178 0.079913107980995587 -
0.29538069766762654 0.39394328174388155 -
0.097795381675424536 0.97680092536560514;-
47.718828012254058 -78.384819237707475 -
20.922777811170167 57.366510664461572
4.1092560946004246 -27.391325703481208 -
21.441766084618628 83.472867999734902
0.33581033109609215 61.900439949174398 -
16.46006980952281 -5.2883855524545655
17.528738567171022;-25.541725833509716
1.619237173206638 9.0973408365424877 -
2.0791070396877793 -1.6571075525831935
7.7738962113921941 -3.018631665568718
4.8360742638482144 -3.2671295018250563
3.4230819328560171 1.2375789286726433 -
2.8379672241416167
9.7650569037510451;1.7427259672558875
0.56555642593638211 1.5019459829113471
0.42296242211352753 -1.3165095534157836
1.4586643689664693 0.27956313228990187
0.33774436316394524 0.38383314782350603
1.3228474840496229 -0.50016771666758519
0.61687652611345956 -3.1972156084233441;-
5.3939244427350133 14.277127528008331 -
15.247185045540963 -4.5023635424427679
18.632301070911673 30.278089939425694
6.8571526378241501 21.791410254600525
40.270442822900407 -61.772106045546735 -
13.178610708585811 7.6935633634734257
33.340303562139688;3.3124594126878431 -
0.09920842851718209 -1.73801863970958
0.70809817398694219 1.1211057707233578 -
1.2558163550695634 0.9070426172621413 -
0.66001347360774243 -2.6681716122608279
2.4343030426294967 -0.6967079798890965
1.4062890184437782 -0.38884529018376163;-
1.2380753697247968 -0.30939007693102072
0.37112729591861893 13.898220360966006
1.9815539353641571 2.4354207199194611 -
0.61340664986142834 1.6309868230079145 -
0.31370159924020102 -0.74581371322647372
0.035648037806396519 1.13008051224665
0.5491079757426125];
% Layer 2 b2 = 0.38351634976819882; LW2_1 = [-0.022979527292961349
0.026871202455676233 0.044731178755804089
-0.71357672644922121 0.008576025304220249
0.02201117689749682 0.2259544252312109
0.0020545652658588518 -0.22490850690782052
0.39085707864842389];
% Output 1 y1_step1_ymin = -1; y1_step1_gain = 0.0444444444444444; y1_step1_xoffset = 5;
% Format Input Arguments isCellX = iscell(X); if ~isCellX, X = {X}; end;
% Dimensions TS = size(X,2); % timesteps if ~isempty(X) Q = size(X{1},2); % samples/series else Q = 0; end
% Allocate Outputs Y = cell(1,TS);
% Time loop for ts=1:TS
% Input 1 Xp1 =
mapminmax_apply(X{1,ts},x1_step1_gain,x1_s
tep1_xoffset,x1_step1_ymin);
% Layer 1 a1 = tansig_apply(repmat(b1,1,Q) +
IW1_1*Xp1);
% Layer 2 a2 = repmat(b2,1,Q) + LW2_1*a1;
% Output 1 Y{1,ts} =
mapminmax_reverse(a2,y1_step1_gain,y1_step
1_xoffset,y1_step1_ymin); end
% Final Delay States Xf = cell(1,0); Af = cell(2,0);
% Format Output Arguments if ~isCellX, Y = cell2mat(Y); end end
% ===== MODULE FUNCTIONS ========
% Map Minimum and Maximum Input Processing
Function function y =
mapminmax_apply(x,settings_gain,settings_x
offset,settings_ymin) y = bsxfun(@minus,x,settings_xoffset); y = bsxfun(@times,y,settings_gain); y = bsxfun(@plus,y,settings_ymin); end
% Sigmoid Symmetric Transfer Function function a = tansig_apply(n) a = 2 ./ (1 + exp(-2*n)) - 1; end
% Map Minimum and Maximum Output Reverse-
Processing Function function x =
mapminmax_reverse(y,settings_gain,settings
_xoffset,settings_ymin) x = bsxfun(@minus,y,settings_ymin); x = bsxfun(@rdivide,x,settings_gain); x = bsxfun(@plus,x,settings_xoffset); end
25 | P a g e
25 Predicting Success of Logistics Network using Artificial Neural Network
5. CONCLUSION
In this brief study of Neural Networks and
its use, we came across how to use the
neural networks for predictions. We also
came across shortest path algorithm
which is used to calculate or traverse the
shortest cost / distance path between two
nodes. We also looked upon various
MATLAB functions and applications which
were used to simulate the neural network,
plot graphs & diagrams, perform data
generation, etc. In this highly competitive
business world, extra cost can be the
difference between a successful and an
unsuccessful business. And, logistics play a
very important role in the fate of a
company as logistics operations are
responsible for movement of resources,
personnel, raw materials, finished goods
and wastages. If the logistics network is
not optimized, then there can be cases of
late arrival, more wastage, more costly
transport, etc.
Sometimes networks can be quite
complex and it becomes tedious to
calculate all possible scenarios and
probabilities. It is where neural networks
are more powerful. Neural networks can
be effectively used to calculate time series
problems, match data clusters, recognize
patterns & solve linear and non-linear
equations fast as well as accurately. The
success of calculations depends hugely on
the correct data so that the network could
easily play around with the analytics. In
case of a neural network, the hidden
layers allow the network more
optimization options.
Fig 14. Neural Network Diagram for the 2nd
example
Now let us discuss a little about the
network as given in the 2nd example. We
can see clearly that there are 13 inputs
given to the network at a particular
instance of time. The hidden layer adapts
itself with each incoming stimuli while
matching it to the targets. In this way,
network training is done. After training is
done, then validation, testing and
simulation of the network takes place. In
validation and testing phase, the neural
network thus achieved from training
phase is tested with different inputs to see
if their output lies within the specified
range. There are different types of neural
network algorithms such as Levenberg-
Marquardt, Resilient Back-propagation,
Scaled Conjugate Gradient, Variable
Learning Rate Back-propagation & many
others.
A neural network is a powerful tool
for the analysis of complex data which can
be sometimes difficult to achieve using
linear programming methods. Finally, we
saw how the neural network was used to
predict an outcome. The output is
achieved in response to an incoming
vector. The prediction value is matched
with the financial values of the company to
arrive at an outcome. One aspect of neural
network which was left out in the previous
sections is that as the number of training
variables increase, so does the quality of
neural network. A more trained neural
network is much more agile and can
predict larger array of variables. There are
similarities between a neural network and
a logistics network, which makes it easy to
analyze the logistics matrix and its
transformation.
Hence the conclusion; that the
neural computing can take analytics and
prediction making to another level. And, if
the neural networks are combined with
some good analytical functions & logic can
make simply outstanding results.
25 | P a g e
25 Predicting Success of Logistics Network using Artificial Neural Network
6. RECOMMENDATIONS
In the last section of this project work, we
would look into what the further work on
this topic could be & the ways to further
improve the network optimization. Since
the neural network which we worked with
depends on historical data so it is
recommended that the data be as original
as possible. The quality of prediction is
directly proportional to the quality of data.
The practical use of Neural
Networks is endless in business situations.
From the viewpoint of a Logistics network,
further experiments can be done with the
neural networks to show the inter-relation
between the different parameters and the
output. In today’s competitive world, it is
also essential to take into account the
quality of service. Using neural networks
we can optimize a path for certain
category of products so that for essential
products the time taken for transportation
is minimum and for non-essential
commodities, the cost is minimum. There
has to be a trade off between the distance
and the cost which could be easily found
by neural computing. Not only so, the
seasonal variations as well as the
randomness of a transportation network
can easily be simulated.
Every company has a Supply Chain
Network for transportation of goods. So,
software designed specifically for logistics
network can be created taking into
account the principles of neural networks.
Not only so, the principles of neural
computing can be applied to other
departments and / or technologies to
pinpoint the growth. Suppose a company
decides to introduce neural networks to
make intelligent predictions about the
manufacturing process, and then what
could be the outcome? Then it would even
be possible that the neural network
automatically read market data ( customer
data, retail data, popularity data, etc ) to
make necessary adjustments to the
product design by itself.
Suppose a neural network
algorithm finds out that customers prefer
books with dimension between 15 cm to
20 cm. Then, the neural network would
automatically instruct the machines to
produce books with the preferred
dimension and the contents inside the
book auto calibrated to fit images and
texts for best possible representation.
Then the neural network would wait to see
the reaction of customers to the newly
designed product and based on the
response it would again modify itself. In
this way, the company would definitely be
ahead of its competitors. Such is the
power and scope of neural networks.
Hence, we saw that the neural
networks have endless usage constrained
by human thoughts and their applications.
Also, newer and more efficient training
algorithms can be designed to enable
further growth in this sector. Coming back
to Logistics network which is a tip of the
iceberg, neural computing can take
business to another level of automation.
26 | P a g e
26 Predicting Success of Logistics Network using Artificial Neural Network
7. REFERENCES
[1] "What is Logistics and Supply Chain Management?", Supplychainopz.com, 2016. [Online].
Available: www.supplychainopz.com/2012/04/what-is-logistics-and-supply-chain-management.html.
[Accessed: 01-Feb- 2016].
[2] "Artificial neural network", Wikipedia, 2016. [Online]. Available:
https://en.wikipedia.org/wiki/Artificial_neural_network. [Accessed: 13-Feb- 2016].
[3] "Stochastic neural network", Wikipedia, 2016. [Online]. Available:
https://en.wikipedia.org/wiki/Stochastic_neural_network. [Accessed: 18- Feb- 2016].
[4] K.A. Smith and J.N. Gupta, Neural networks in business: techniques and applications for the operations
researcher, 1st ed. PERGAMON, 2000.
[5] M. Kazemi, A. Niknafs, V. Ranjbar and A. Forouharfar, APPLICATION OF NEURAL NETWORKS IN
FORECASTING BUSINESS AND MANAGERIAL PROCESSES IN COMPARISON WITH NONLINEAR MODELS
(CASE STUDY: IRAN’S WOOD INDUSTRY), 1st ed. International Journal of Social and Economic Research,
2011, pp. 220-225.
[6] S. Rajamanickam, A. Maithili and R. Kumari, NEURAL NETWORK TOWARDS BUSINESS FORECASTING, 2nd
ed. IOSR Journal of Engineering, 2012, pp. 831-836.
[7] Y. Eldon, Artificial neural networks and their business applications, 1st ed. Taiwan: ELSEVIER, 1994, pp.
303-313.
[8] A. Moutinho and J. Costa Sousa, "NETWORK OPTIMIZATION MODELS", Presentation, 2016.
[9] S. Sivanandam, S. Sumathi and S. Deepa, Introduction to neural networks using MATLAB 6.0. New Delhi:
McGraw Hill Education (India) Private Limited, 2006.
[10] E. Davalo and P. Naïm, Neural networks. London: Macmillan, 1991.
[11 ] G. Swamy and G. Kumar, Neural Networks. Hyderabad, India: SCITECH, 2016.
[12] V. Sharma, Artificial Neural Network Applicability in Business Forecasting, 1st ed. Jammu: International
Journal of Emerging Research in Management &Technology, 2012.
27 | P a g e
27 Predicting Success of Logistics Network using Artificial Neural Network
8. APPENDIX
( List of Tables & Figures used in the project work )
List of Figures
1. Artificial Neural Network Model ( Pg- 2 )
2. Probability Density Function ( Pg – 3 )
3. Schematic of a Neuron ( Pg – 4 )
4. Abbreviated Logistics / Supply Chain Network ( Pg – 7 )
5. Nodal points of a Logistics / Supply Chain Network ( Pg – 7 )
6. MSE graph for 1st Training period ( Pg – 18 )
7. MSE graph for 2nd Training period ( Pg – 18 )
8. Training State of ANN ( Pg – 18 )
9. NARX Neural Network Diagram ( Pg – 18 )
10. Regression Plot for the Network ( Pg – 19 )
11. Performance Graph for 2nd example ( Pg – 21 )
12. Gradient, mu and validation fail plot for 2nd example ( Pg – 21 )
13. Regression plot for 2nd example ( Pg – 22 )
14. Neural Network Diagram for the 2nd example ( Pg – 24 )
List of Tables
1. Cost Matrix of the Network ( in INR Lakhs ) ( Pg – 10 )
2. Distance Matrix of the Network ( in Kilometres ) ( Pg – 10 )
3. Probability / Noise Matrix ( Pg – 11 )
4. Actual Cost Matrix ( in INR Lakhs ) ( Pg – 11 )
List of Neural Network Coding
1. Code 1 ( Pg – 15 )
• Network based on a logistics network which takes in a network matrix as input and
the shortest path distance between the nodes as output.
2. Code 2 ( Pg – 22 )
• Network based on giving 13 parameters as input and getting the total cost as
output. The network is also used to predict the outcomes by providing a 13*1
matrix as test input which represents the 13 parameters.