Outline

1

• PROJECT PROPOSAL

• Data Classification using Neuro-Fuzzy Approach
  (NEFCLASS Model)
• CSCI 6501 Intelligent Systems
• Group Members
• Tayyaba Sharif
• Ling Ou
• Chen Yang
• Rafiy Saleh

2
Outline

• Project Goal
• Background
• Papers review
• Conclusion
• References

3
Project Goal

• What we are going to do in this project?
• To develop a neuro-fuzzy system for the
  classification of data
• The system is based on a generic model of a
  fuzzy perceptron.
• Our Approach
• Train neural network using training data set
• Training data (can not overlap each category
  and no fuzziness) ? rules (fuzzy error back
  propagation) ? classify data

4
The Data Set

• Iris data is perhaps the best known database to
  be found in the pattern recognition literature.
• Detail of Iris data set
• Use to classify three types of Iris flowers by
  the length and width of their sepals and petals.
• Data set contains 3 classes, 4 numeric predictive
  attributes which describe the sepal and petal of
  the Iris, and 150 instances.
• Data set is split in half and the patterns are
  ordered alternately within the training and test
  data sets

5
Background

• Classification
• Neural-network
• Fuzzy logic
• Neuro-fuzzy
• Advantages of this approach

6
Architecture of the system
Layer 1 Crisp inputs
Layer 2 Input membership functions
Layer 3 Fuzzy rules
Layer 4 Output membership functions
x1
R1
A1
x1
x1
R2
A2
x1
C1
R3
A3
x2
R4
B1
x2
C2
x2
R5
B2
x2
R6
B3
7
Architecture of the system

• The first layer is the input variables that means
  the pattern instances,
• The hidden layer represents fuzzy rules,
• The third layer stores the output variables that
  means one unit for each class,
• The units use t-norms and t-conorms as activation
  functions,
• t-norm, t0,1x0,1-gt0,1, t(x,y)min(x,y)
• its also called fuzzy intersection.
• t-conorm , t0,1x0,1-gt0,1, t(x,y)max(x,y)
• its also called fuzzy union
• The fuzzy sets are encoded as (fuzzy) connection
  weights.

8
A fuzzy perceptron as a generic model for
neuro-fuzzy approaches, by Detlef Nauck

• Fuzzy perceptron
• Provides an architecture that can be initialized
  with the prior knowledge for training the neural
  networks
• Main features of Fuzzy perceptron.
• They are specialized for data analysis and
  control task
• The Architecture of Fuzzy percepton is same as
  multilayer perceptron. The only difference is
  that the weights of the connections are modeled
  as fuzzy sets, which changes the activation,
  output and propagation functions accordingly.
• Fuzzy perceptron put prior knowledge into the
  neural network shortening the learning process.

9
Main features of Fuzzy perceptron (Contd)

• It provides mechanism to deduce the learned
  result in form of linguistic rules.
• Fuzzy perceptrons are build with the odd number
  of hidden layers. In his paper Detlef Nauck
  discusses about 3 layer feedforward neural
  network (i.e. the neural network with one hidden
  layer).
• It commonly uses triangular functions as
  membership functions.Triagular membership
  functions can be defined as
• Triangle(xa,b,c)maxmin(x-a)/(b-a),(c-x)/(c-b
  ),0
• It uses back propagation algorithm for adjusting
  the weights.

10
NEFCON

• It is a model used for neuro-fuzzy control and is
  derived from generic fuzzy perceptron.
• Features of NEFCON
• It is a special 3 layer fuzzy perceptron
• The connections between neurons are represented
  by linguistic terms
• Its learning algorithm uses a special version of
  fuzzy error back propagation
• It uses reinforcement learning which allows
  adaptation of membership function online.

11
NEFCLASS A Neuro-Fuzzy Approach for the
Classification of Data by Detlef Nauck, Rudolf
Kruse

• Main Features
• The NEFCLASS model is a neuro-fuzzy
  classification system derived from a generic
  model of a 3-layer fuzzy perceptron
• For each linguistic value there is only one
  representation as a fuzzy set.
• It shares the weights between the connections and
  the connections which share the weights come from
  the same input unit.
• It uses same triangular membership function as
  NEFCON.

12
The NEFCLASS Algorithm

• The learning process (training of the network)
  ?weights are changed such that in the end a
  classifier is created.
• 2 Learning process steps
• First, the structure of the classifier that fits
  the respective database is created
• Then the classifier is completed by determining
  the parameters of the system in an iterative
  training process to improve the accuracy without
  loosing the interpretability.

13
The NEFCLASS Algorithm

• 1st Learninig step Learning a Rule Base
• A NEFCLASS system is built from partial knowledge
  about the patterns
• For each input variable the user decides the
  number of fuzzy sets to be used
• Define the membership function?Trianglar
  membership function
• Initialize NEFCLASS with k lt kmax fuzzy rules
• The rule base of NEFCLASS is completed by finding
  for each pattern a combination of fuzzy sets that
  yields the highest degree of membership for each
  value of pattern.
• After rule learning NEFCLASS can do
  classification but it need improvement so 2nd
  learning step is required.

14
The NEFCLASS Algorithm

• 2nd learninng step Training Fuzzy Sets
• NEFCLASS uses supervised learning algorithm to
  adapt the fuzzy sets until admissible error or
  misclassification reached.
• Propagate the pattern
• Determine error for each output
• Error decides ,how the changes in the membership
  function should be made using a simple heuristic
• fuzzy set is changed accordingly by altering the
  parameters of its membership function.

15
The NEFCLASS Algorithm

• Error decides ,how the changes in the membership
  function should be made using a simple heuristic
• fuzzy set is changed accordingly by altering the
  parameters of its membership function.

16
Evaluation of the System

• Goal
• Compare the validity of the classifier and other
  properties in order to evaluate the NEFCLASS
  model
• Based on the same Iris data set, the paper
  compares the NEFCLASS model with
• a common multilayer perceptron
• another neuro-fuzzy approach, FuNe I

17
NEFCLASS V.S. Fuzzy Percetron
18
NEFCLASS V.S. FuNe I
19
Summary of Evaluation

• NEFCLASS, a generic model of a 3-layer fuzzy
  percetron, and another neuro-fuzzy approach, FuNe
  I, are comparable in their performances.
• NEFCLASS is easy to implement, easy to handle and
  easy to understand.
• The learning and pruning strategies of NEFCLASS
  are simple and fast heuristics.

20
Conclusion

• The neuro-fuzzy systems can be used for data
  mining especially the classification (development
  of decision support systems) of medical
  ,engineering or business field.

21
Reference

• 1 Detlef Nauck, Rudolf Kruse, NEFCLASS A
  Neuro-Fuzzy Approach for the Classification of
  Data, Symposium on Applied Computing, Nashville,
  1995.
• 2 Detlef Nauck, A Fuzzy Perceptron as a
  Generic Model for Neuro-Fuzzy Approaches, In
  Proc. Fuzzy Systeme94, Munich, 1994.

22
(No Transcript)