A Self-Organized Network inspired by Immune Algorithm

1
A Self-Organized Network inspired by Immune
Algorithm
2002?? ?????? ????

• M. Rahmat WIDYANTO (D1)
• Hirota Laboratory
• Computational Intelligence Systems Science
• Tokyo Institute of Technology

2
Contents

• Theoretical Section
• Application Section
• A Time-Temperature-based Food Quality Prediction
  using a Self-Organized Network inspired by Immune
  Algorithm
• R. Widyanto, Megawati, Y. Takama, K. Hirota
• (to be submitted to International Conference on
  Soft Computing and Intelligent System 2002,
  Tsukuba, Japan)
• A Self-Organized Network inspired by Immune
  Algorithm for Clustering Analysis
• R. Widyanto, Megawati, K. Hirota
• (to be submitted to The 2002 IEEE International
  Conference on Data Mining, Maebashi City, Japan)
• Generalization Improvement of Prostate Cancer
  Prediction using a Self-Organized Network
  inspired by Immune Algorithm (Future Work)

3
Theoretical Section Background

• Self-Organized Network Kohonen, 1996
• Number of neurons should be decided
• Describes characteristics from trained data only
• Low Generalization Ability

• Immune Algorithm Timmis, 2001
• First and Second Immune Responses
• Automatic Creation of B-cells
• Mutation of B-cells

• Improved Version of Self-Organized Network
• Neurons are
• automatically created
• Generalization is
• improved

4
Application Section (1)Prediction System
5
Application Section (1)Data Acquisition

• Akita is known as a pork production area.
• Everyday frozen trucks deliver the meat from
  Akita to Chiba.
• During delivery, temperature inside the trucks is
  recorded every 5 minutes using data lodger.
• Data lodger consists of two channels.
• Channel 1 to measure the meat packaging box.
• Channel 2 to measure the meats surface.

6
Application Section (1)Neural Network

• Combined with back-propagation output layer

7
Application Section (1)Experiment Setting

• For each region TOP, MIDDLE, BOTTOM, experiment
  is conducted separately.
• From October to December 2001 there were 15 times
  meat deliveries from Akita to Chiba resulting 15
  input data.
• Recognition Experiment
• Learning Phase all 15 data trained to network
• Testing Phase all 15 data tested
• Compare recognition obtained by SONIA network and
  standard back-propagation network.
• The codes are implemented on PC (600 Mhz
  Processor, 64 MB RAM) using Matlab 6.1 under
  Windows 2000 operating system.

8
Application Section (1)Experiment Error
Convergence
BOTTOM
TOP
MIDDLE

• Back-propagation slower convergence (red-line)
• SONIA network faster convergence (green-line)

9
Application Section (1)Experiment Recognition
Result

• SONIA network outperformed back-propagation in
  recognition experiment.

10
Application Section (1)Experiment Computation
Time

• SONIA network slightly needed more computation
  time in average than back-propagation network. (s
  seconds)