Project Progress Report

1
Project Progress Report
Data Classification using Neuro-Fuzzy Approach
Presented by Yang Chen
Rafiy Saleh Ling Ou
Tayyaba Sharif
2
Projects Goal

• To realize a hybrid neuro-fuzzy model for
  classification.
• We adopt Habermans Survival Data as our
  experiment data set. After number of training
  epochs, the system should tell whether or not a
  patient can live more than 5 years according to
  three input parameters.

3
Motivation

• Neural networks are low-level computational
  structures that perform well when dealing with
  raw data, however they are opaque to user.
• Fuzzy logic deals with reasoning on a higher
  level, using linguistic information acquired from
  domain expert, however it lacks the ability to
  learn and can not adjust itself to a new
  environment.
• Our approach combines the advantages of those two
  approaches.
• Training part to develop IF-THEN fuzzy rules, and
  determine membership functions for input and
  output variables of the system. Expert knowledge
  can be easily incorporated into the structure of
  the system.

4
Main parts of our system

• Use a matrix to represent the relationship of the
  nodes in our neuro-fuzzy network.
• Generate the initial parameters for layer 1 and
  layer 4.
• Use a hybrid neuro-fuzzy adaptive networks to
  train the parameters for layer 1 and layer 4.
• After the parameters of the network have been
  decided, the actions of the network will be
  decided. We can use test the performance of our
  classifier.

5
Architecture of the Adaptive Network

• Adaptive Network
• A multi-layer feedforward network
• Each node
• performs a node function on incoming signals and
  a set of parameters
• Either square node (adaptive node)
• Has parameters
• Or circle node (fixed node)
• No parameters
• Links
• Indicate flow direction of signals between nodes
• Has no weights associated with

6
Network Structure
Consequent Parameters
Premise Parameters
1
1
1
A1
x
2
A2
2
2
3
A3
3
3
B1
?
f
B2
y
B3
25
25
25
C1
26
26
26
z
C2
27
27
C3
27
7
The Network Structure

• Has 3 input variables with 27 rules
• Three membership functions are associated with
  each input
• So the input space is partitioned into 27 fuzzy
  subspaces each of which is governed by a fuzzy
  IF-TEHN Rule.
• The premise part of a rule defines a fuzzy
  subspace and the consequent part specifies the
  output within this subspace
• x, y, z are inputs to nodes Ai, Bi, Ci(Linguistic
  labels)

8
Architecture of the Adaptive Network

• Layer 0
• Layer 1
• Nodes are square nodes with node functions
• Oi 1 µAi (x)
• Oi 1 µBi (x)
• Oi 1 µCi (x)
• For Oi 1 µAi (x),
• Input to the node i x
• Linguistic label associated with this node
  function Ai
• Membership function of Ai Oi 1
• Bell function µAi (x)

9
Membership Function
To get a bell-shaped membership distribution with
maximum equal to 1 and minimum equal to 0 we use
this membership function
Where ai, bi, ci is the premise parameter set.
The values of these parameters change and the
bell-shaped functions vary accordingly, this
exhibit various form of membership functions on
linguist label Ai
10
Architecture of the Adaptive Network

• Layer 2
• Nodes are circle nodes with outputs presenting
  the firing strength of rules.
• wi µAi (x) µBi (y) µCi (z), i 1, 2,
  3
• Layer 3
• Nodes are circle nodes.
• For ith node, the ratio of ith rules firing
  strength to the sum of all rules firing
  strengths
• wi wi / (w1 w2 w3), i 1, 2, 3

11
Architecture of the Adaptive Network

• Layer 4
• Nodes are square nodes with node function
• Oi4 wi fi
• wi (pix qiy riz
  ji)
• Output of layer 3 wi
• Consequent parameters pi ,qi , ri , ji
• Layer 5
• Single circle node in this layer with overall
  output as the summation of all incoming signals
• O15 overall output
• ? wi fi
• ? wi fi / ? wi

12
Basic steps for part 3

• Input the command and parameters according to the
  training dataset
• Create data structures according to those
  parameters
• Read the nodes structure generated in part 1
• Build a neuro-fuzzy network
• Read the initials parameters generated in part2
• Read the training data
• For (I 0 I lt epoch number i)
• For (j 0 j lt dataset number j)
• Read the input
• Calculate the nodes output of layer 1,
  layer 2, and layer 3
• Put the result into a 2-dimensional data
  structure
• Put a kalman filter to the data structure to
  generate karman filter
• Use least squares method to generate the
  parameters for layer 4
• For (j 0 j lt dataset number j)
• Calculate the output of layer 4 and layer
  5
• Calculate the error
• Use gradient descent method to update the
  parameters for layer 1

13

• CURRENT STATUS OF THE PROJECT

14
Applying the model to our haberman dataset

• 1. dataset

15

• 2. Initial parameter generate by the system

16

• 3. Matrix, standing for the node connection of
  the neuro-fuzzy network, generated by our system.

..
A matrix with 94 rows and 94 columns
17

• 4. Final parameters for the Neuro-Fuzzy network

18
Fuzzy set for age
Fuzzy set generated using the parameters
generated in step 4
19
Fuzzy set for year of operation
20
Fuzzy set for number of nodes
21
Conclusion
22
References

• Haberman data set,
• ftp//ftp.ics.uci.edu/pub/machine-learning-databas
  es/
• Adaptive-Network-Based Fuzzy Inference System
  http//wwwmath.uni-muenster.de/SoftComputing
• /lehre/seminar/ss2002/jang93anfis.pdf