Transcript شبکه های عصبی - University of Tehran

‫شبکه های عصبی‬
‫ارائه کنندگان ‪:‬‬
‫بنیامین قاسمی‬
‫ساره منطقی‬
‫فهرست مطالب‬
‫معرفی شبکه عصبی‬
‫معایب و مزایا‬
‫شبکه عصبی مصنوعی‬
‫استفاده در مدیریت‬
‫انواع شبکه های عصبی‬
‫مثال ارائه شده‬
‫نحوه ی کارکرد‬
‫جمع بندی مطالب‬
‫‪End‬‬
‫معرفی شبکه های عصبی‬
What is a Neural Network?
 Collection of “neurons”
 Computes some function
 Takes input
 Produces output
 Can learn
‫معرفی شبکه های عصبی‬
Human Brain Function
Human brain can generalize from abstract
Recognize patterns in the presence of noise
Recall memories
Make decisions for current problems based on prior experience




Neural Network Neurons




Receives n-inputs
Multiplies each input by its weight
Applies activation function to the sum of
results
Outputs result
‫شبکه عصبی مصنوعی‬
What is an Artificial
Neural Network (ANN)?
The computational ability
of a digital computer
combined with the
desirable functions of the
human brain.
‫شبکه عصبی مصنوعی‬
Biological Neurons
Artificial Neurons
http://research.yale.edu/ysm/images/78.2/articles-neural-neuron.jpg
http://faculty.washington.edu/chudler/color/pic1an.gif
‫شبکه عصبی مصنوعی‬
‫ساختار شبکه عصبی مصنوعی‬
Input Units
Influence
Map
Layer 1
Influence
Map
Layer 2
Hidden Units
Output Units
‫شبکه عصبی مصنوعی‬
‫تناظر بین شبکه عصبی و شبکه مصنوعی‬
Biological Neural Network
Soma
Dendrite
Axon
Synapse
Artificial Neural Network
Neuron
Input
Output
Weight
Soma
Synapse
Dendrites
Axon
Soma
Input Signals
Axon
Out put Signals
Synapse
Dendrites
Middle Layer
Synapse
Input Layer
Output Layer
‫نحوه ی کارکرد‬
How does a neural network learn?




A neural network learns by determining the
relation between the inputs and outputs.
By calculating the relative importance of the
inputs and outputs the system can determine
such relationships.
Through trial and error the system compares its results with the
expert provided results in the data until it has reached an
accuracy level defined by the user.
With each trial the weight assigned to the inputs is changed until
the desired results are reached.
‫نحوه ی کارکرد‬
How the Process Works ?
Step 1: Initialisation

Set initial weights to random numbers in your
range
Step 2: Activation

Activate the perceptron by applying inputs
and desired output.( training set data)
Calculate the actual output at iteration .
‫نحوه ی کارکرد‬
Step 3: Weight training

Update the weights of the perceptron.The weight
correction is computed by the delta rule
Step 4: Iteration

Increase iteration p by one, go back to Step 2 and
repeat the process until convergence.
‫انواع شبکه های عصبی‬
Types of Networks:




Multi-Layer-Perceptron
Hopfield Net
Kohonen Feature Map
Adaptive Resonance
Theory (Art), Fussy
ArtMap
Different Learning
Algorithms:






Type of Learning:


Supervised
Unsupervised
Backpropagation
Delta Learning Rule
Forward Propagation
Hebb Learning Rule
Simulated Annealing
Genetic Algorithms
‫انواع شبکه های عصبی‬
Muli-Layer Perceptron
Hopfield Net
‫معایب و مزایا استفاده از شبکه ها‬
Neural Networks can be
extremely complex and
hard to use
The programs are filled
with settings you must
input and a small amount
of data will cause your
predictions to have error
The results can be very
hard to interpret as well
Dead-end situations are
hard to avoid
Neural networks can find
relations that no one
ever guess they exist
Since they are data
dependent performance
will improve as sample
size increases
Regression performs
better when theory or
experience indicates an
underlying relationship
‫استفاده در مدیریت‬
Marketing
Trading and financial forecast
Future price estimation
Exchange rate forcast
Bankruptcy prediction
Stock performance and selection
Portfolio assignment and optimization
‫مثال کاربردی‬
Large amount of data available in
databases
Customers data available in firm’s own
database or can be supplied by
companies which sell these information
These information can be applied for
marketing purposes e.g. direct marketing
‫مثال کاربردی‬
Direct marketing drives high cost
Targeting customers who are more likely
to spend money !
Direct mailing to customers
‫مثال کاربردی‬
In this example one charity organization apply
direct mailing promotion to raise their funds
Neural network applied to target selection in this
case
Neural network should determine those
customers in data base who would be interested
in the offer being maid
Neural network in a learning system which can
adapt the nonlinearity in the data to capture
complex
‫مثال کاربردی‬
There can be different types of databaises
with variety of data
There should be most important aspects of
a successful mailing compa

Analytical methods (data mining, sensitivity analysis, … )
Experiment ( some researches , literature , …)
Causal relations

experties


‫مثال کاربردی‬
Data mining offers following
representations as purchase history:
1.
2.
3.
Recency of purchase
Frequency of purchase
Monetary value
These variables are called RFM
variables
‫مثال کاربردی‬
In working with models like ANN enough care
must be taken about the process & the data
data preparation
publish in–publish out
Determining causal relations
Knowledge about customer’s attitude & history
‫مثال کاربردی‬
Mailing strategy
Who should be mailed and how frequent
How frequently ** Should be organized
How their promotional material should be
organized
‫مثال کاربردی‬
Classification or prediction
Asymmetrical misclassification costs
Weigh misclassified responders
Or target scoring to show customers
willingness
‫مثال کاربردی‬
What is ANN job
Trained to determine correct set of
network parameters
Good indication of the willingness
according to network inputs
Indeed, indication of responsive behavior
regarding their characteristics
A nonlinear regression model
‫مثال کاربردی‬
Network configuration
feed-forward neural networks for practical
purpases
number of hidden layers
methods: growing and pruning, heuristic
search, optimization by evllutionary
computation (e.g. GA). experiment,...
‫مثال کاربردی‬
selecting network parameters
experiments show that one hidden layer
provides model with sufficient accuracy in
target selecting
transfer function: hyperbolic tangent
sigmoid or logistic sigmoid,...
‫مثال کاربردی‬
data preparetoin
discription of raw data
size of data set
feature selection
process selected features
selecting suitabale training and validation
sets
‫مثال کاربردی‬
description of raw data
a well-known Dutch charuty organization
more than 725000 supporters in the internal
database
database including:



mailing dates,
amount of donation,
date of donation in response to a particular mailing,...
‫مثال کاربردی‬
size of data set
aoge amount of records makes network
too complex and slow
data size should be large enough and not
too big
1000 random records representativies of
the whole data
‫مثال کاربردی‬
Feature selection
neccessarily not all features are useful and
meaningful for target selection
features that sumarize the most important
aspects
RFM variables
table 2: the features used for the charity case study
‫جمع بندی مطالب‬
Neural networks provide ability to provide
more human-like AI
Takes rough approximation and hardcoded reactions out of AI design (i.e.
Rules and FSMs)
Still require a lot of fine-tuning during
development
‫منابع‬
Artificial Neural Network in Finance and Manufacturing,Joarder
Kamruzzaman
Neural Networks And Their Statistical Application, Clint Hagen
Statistics Senior Seminar 2006
Neural Networks , Megan Vasta
Artificial Intelligence & Neural Networks, Amir Hesami
Interview with Jeff Hannan, creator of AI for Colin McRae Rally 2.0
Interview with Derek Smart, creator of AI for Battlecruiser: 3000AD
Neural Netware, a tutorial on neural networks
Sweetser, Penny. “Strategic Decision-Making with Neural Networks and
Influence Maps”, AI Game Programming Wisdom 2, Section 7.7 (439 – 46)
Russell, Stuart and Norvig, Peter. Artificial Intelligence: A Modern Approach,
Section 20.5 (736 – 48)
‫از توجه شما بسیار سپاسگزاریم‬
‫ساره منطقی‬
‫بنیامین قاسمی‬
‫پاییز ‪1386‬‬