dept. of computational intelligence & systems science tokyo institute of technology

56
1/56 Family of Family of S S elf- elf- O O rganized rganized N N etwork etwork Inspired by Inspired by I I mmune mmune A A lgorithm lgorithm ( ( SONIA SONIA ) and ) and Their Various Applications Their Various Applications Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology Muhammad R. Widyanto 03D35190 博博博博博 博博博博博 2006.01.04 2006.01.04 SONIA F-SONIA SONIA-DNN CMF-SONIA EF-SONIA

Upload: iria

Post on 31-Jan-2016

59 views

Category:

Documents


0 download

DESCRIPTION

博士本審査 2006.01.04. Family of S elf- O rganized N etwork Inspired by I mmune A lgorithm ( SONIA ) and Their Various Applications. Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology. SONIA. SONIA-DNN. CMF-SONIA. Muhammad R. Widyanto 03D35190. F-SONIA. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

1/56

Family of Family of SSelf-elf-OOrganized rganized NNetwork Inspired etwork Inspired by by IImmune mmune AAlgorithm (lgorithm (SONIASONIA) and ) and

Their Various ApplicationsTheir Various Applications

Dept. of Computational Intelligence & Systems ScienceTokyo Institute of Technology

Muhammad R. Widyanto03D35190

博士本審査 博士本審査 2006.01.042006.01.04

SONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

Page 2: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

2/56

Thesis Road MapThesis Road MapChapter 1Chapter 1

IntroductionIntroduction

Chapter 2 Chapter 2 [J1][J1]SONIA and Food Quality PredictionSONIA and Food Quality Prediction

[J[Jxx]: Journal Paper ]: Journal Paper xx-th-th

Chapter 3 Chapter 3 [J2][J2]SONIA-DNN and Preference ModelingSONIA-DNN and Preference Modeling

Chapter 4 Chapter 4 [J3][J3]F-SONIA and Fragrance RecognitionF-SONIA and Fragrance Recognition

Chapter 5 Chapter 5 [J4][J4]CMF-SONIA and Overlapping Pat. Clas.CMF-SONIA and Overlapping Pat. Clas.

Chapter 6 Chapter 6 [J5][J5]EF-SONIA and Unknown Odor Recog.EF-SONIA and Unknown Odor Recog.

Chapter 7Chapter 7ConclusionsConclusions

SONIA

F-SONIASONIA-DNN CMF-SONIA

EF-SONIA

Page 3: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

3/56

ContentsContentsSONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

Chap. 3 SONIA-DNN for Preference ModelingChap. 3 SONIA-DNN for Preference Modeling

Chap. 4 F-SONIA for Fragrance RecognitionChap. 4 F-SONIA for Fragrance Recognition

Chap. 5 CMF-SONIA for Overlapping Pattern Class.Chap. 5 CMF-SONIA for Overlapping Pattern Class.

Chap. 6 EF-SONIA for Unknown Odor RecognitionChap. 6 EF-SONIA for Unknown Odor Recognition

Chap. 7 ConclusionsChap. 7 Conclusions

Chap. 2 SONIA and Food Quality PredictionChap. 2 SONIA and Food Quality Prediction

Chap. 1 IntroductionChap. 1 Introduction

Page 4: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

4/56

Problems Problems Chap. 1 IntroductionChap. 1 Introduction

Global ResponseGlobal Response

OverfittingOverfitting

Low GeneralizationLow Generalization

BPNN BPNN [Rumelhart, 86][Rumelhart, 86]Back-Propagation Neural NetworkBack-Propagation Neural Network

Page 5: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

5/56

OpportunityOpportunityChap. 1 IntroductionChap. 1 Introduction

Immune Algorithm Immune Algorithm [Timmis, 01][Timmis, 01]

Local ResponseLocal Response

Characteristics only Characteristics only

Diverse RepresentationDiverse Representation

Page 6: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

6/56

Chap. 1 IntroductionChap. 1 Introduction

SolutionSolution

BPNN BPNN [Rumelhart,86][Rumelhart,86] Immune AlgorithmImmune Algorithm [Timmis,01][Timmis,01]

Better RecognitionBetter Recognition

Better Generalization Better Generalization

A A SSelf-elf-OOrganized rganized NNetwork etwork inspired by inspired by IImmune mmune AAlgorithmlgorithm[proposed][proposed]

SONIASONIA

Page 7: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

7/56

Chap. 1 IntroductionChap. 1 IntroductionApplicationsApplications

SONIASONIA

Food Quality PredictionFood Quality Prediction

Preference ModelingPreference Modeling

Fragrance RecognitionFragrance Recognition

Unknwon Odor Recog.Unknwon Odor Recog.

Overlapping Pat. Clas.Overlapping Pat. Clas.

SONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

Page 8: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

8/56

Chap. 1 IntroductionChap. 1 Introduction

ContentsContents

Chap. 2 SONIA and Food Quality PredictionChap. 2 SONIA and Food Quality Prediction

Chap. 3 SONIA-DNN for Preference ModelingChap. 3 SONIA-DNN for Preference Modeling

Chap. 4 F-SONIA for Fragrance RecognitionChap. 4 F-SONIA for Fragrance Recognition

Chap. 5 CMF-SONIA for Overlapping Pattern Class.Chap. 5 CMF-SONIA for Overlapping Pattern Class.

Chap. 6 EF-SONIA for Unknown Odor RecognitionChap. 6 EF-SONIA for Unknown Odor Recognition

Chap. 7 ConclusionsChap. 7 Conclusions

SONIA

Page 9: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

9/56

・・・

・・・

・・・

Input layer

Hiddenlayer

Outputlayer

BPNN :BPNN :[Rumelhart,86]

Input Vector Hidden Unit

Antigen

Immune Immune Algorithm :Algorithm :[Timmis,01]

Recognition Ball (RB)

SSelf-elf-OOrganized rganized NNetwork inspired by etwork inspired by IImmune mmune AAlgorithmlgorithm

[proposed]

Chap. 2 SONIAChap. 2 SONIA

Page 10: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

10/56

Input Vector

Unit Centroid

Hidden UnitHidden UnitRecognition Ball (RB)Recognition Ball (RB)

B CellAntibody

Antigen

Paratope

Epitope EuclidianDistance

Chap. 2 SONIAChap. 2 SONIA

Recognition Ball & Hidden Unit Recognition Ball & Hidden Unit [proposed]

Page 11: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

11/56

Antibody GenerationAntibody Generation [Timmis,01][Timmis,01]

Antigen[1..m]

Input Vector [1..m]

Hidden Unit CreationHidden Unit Creationof BPNN of BPNN [proposed][proposed]

Hidden Unit 1

Hidden Unit 2

Hidden Unit i

MutatedHidden Unit n

RB 2

RB i

Mutated RB n

RB 1

B-Cell Construction & Mutation B-Cell Construction & Mutation Chap. 2 SONIAChap. 2 SONIA

[proposed]

Page 12: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

12/56

Training Data

Approximation

Chap. 2 SONIAChap. 2 SONIABPNN BPNN [Rumelhart, 86][Rumelhart, 86]

Approximation Error : 0.01994

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

x

h(x)

BPNN Regularization BPNN Regularization [MacKay, 92][MacKay, 92]

Approximation Error : 0.00241

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

x

h(x)

SONIA without mutation SONIA without mutation Approximation Error : 0.01008

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

x

h(x)

SONIA with mutation SONIA with mutation [proposed][proposed]

Approximation Error : 0.00118

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

x

h(x)

Page 13: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

13/56

Supermarket

Food StoreMarket Area

Production Area

Frozen Truck

Perishable Food

Quality Control Server

Prediction Engine:Neural Networks

Quality Quality CheckCheck

Chap. 2 SONIAChap. 2 SONIA

Food Quality PredictionFood Quality Prediction

Collaborative Project with Japan Ministry of Agriculture and CSD Inc.

Page 14: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

14/56

Data Collection :Data Lodger

Data Collection :Data Lodger

011016_t

- 5051015202530

1 23

45

67

89

111

133

155

177

199

Series1Series2

Channel 1

Channel 2

Time-temperature Data

Time

oC( X 5 Minutes )

Feature Extraction :Mean & Standard Deviation

Feature Extraction :Mean & Standard Deviation

Range Selected

A B C D E

Neural Networks

ch1:Mean

ch1:SD

ch2:Mean

ch2:SD

Quality

good

Pre-Processing :Range Selection

Pre-Processing :Range Selection

Chap. 2 SONIAChap. 2 SONIA

Prediction SystemPrediction System [proposed]

Collaborative Project with Japan Ministry of Agriculture and CSD Inc.

Page 15: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

15/56

Chap. 2 SONIAChap. 2 SONIARecognition AccuracyRecognition AccuracyCollaborative Project with Japan Ministry of Agriculture and CSD Inc.

TOP

Recognition (%)

100

50

0MIDDLE BOTTOM

TOP

MIDDLE

BOTTOM

SONIA

BPNN

Page 16: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

16/56

Chap. 1 IntroductionChap. 1 Introduction

ContentsContents

Chap. 2 SONIA and Food Quality PredictionChap. 2 SONIA and Food Quality Prediction

Chap. 3 SONIA-DNN for Preference ModelingChap. 3 SONIA-DNN for Preference Modeling

Chap. 4 F-SONIA for Fragrance RecognitionChap. 4 F-SONIA for Fragrance Recognition

Chap. 5 CMF-SONIA for Overlapping Pattern Class.Chap. 5 CMF-SONIA for Overlapping Pattern Class.

Chap. 6 EF-SONIA for Unknown Odor RecognitionChap. 6 EF-SONIA for Unknown Odor Recognition

Chap. 7 ConclusionsChap. 7 Conclusions

SONIA-DNN

Page 17: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

17/56

DDecision ecision MMaker (aker (DMDM) Preference) Preference

Price: 5 million yenPrice: 5 million yen

Engine: 3000 ccEngine: 3000 cc

Consumption: 10km/lConsumption: 10km/l

PreferencePreferenceValueValue

DDecision ecision MMaker (aker (DMDM))

Alternative1:Alternative1: Nissan FugaNissan Fuga

Modeling Modeling DM Preference ???DM Preference ???

Chap. 3 SONIA-DNNChap. 3 SONIA-DNN

JSPS Center Of Excellence Project

Page 18: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

18/56

Preference Value by ComparisonsPreference Value by Comparisons

Alternative1Alternative1:: Nissan FugaNissan Fuga

Alternative2:Alternative2: Toyota Mark XToyota Mark X

ComparisonComparisonValueValue

DDecision ecision MMaker (aker (DMDM))

Chap. 3 SONIA-DNNChap. 3 SONIA-DNN

Page 19: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

19/56

ComparisonComparisonValueValue

SONIA(1)

SONIA(2)

SONIASONIA-based -based DDecision ecision NNeural eural NNetwork etwork [proposed]

Alternative 1

Alternative 2

Chap. 3 SONIA-DNNChap. 3 SONIA-DNN

JSPS Center Of Excellence Project

Incomplete Comparisons

Better Generalization

Page 20: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

20/56

Alter-native 1

Alter-native 2

Alter-native 3

・ ・ ・

Alter-native n

Alter-native 1

ー 1.2 0.8 1.3 0.9

Alter-native 2

ー ー 1.1

Alter-native 3

ー ー ー 0.7・・・

ー ー ー ー

Alter-native n

ー ー ー ー ーDDecision ecision

MMaker (aker (DMDM)) Limited Training DataLimited Training Data

Too many!

Incomplete ComparisonsIncomplete ComparisonsChap. 3 SONIA-DNNChap. 3 SONIA-DNN

JSPS Center Of Excellence Project

Page 21: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

21/56

Lp-metric Function BenchmarkLp-metric Function Benchmark [Sun, 1996]

AlternativeAlternativePreferencePreference

ValueValue

DDecision ecision MMaker (aker (DMDM))

1/

*

1

( )pI

pi i i

i

V L z z

*

: Preference Value, : Maximum Value

: Number of Criteria, : Weight Parameter

: Max Vector Value, : Alternative Vector

: Number of Dimension

V L

I

z z

p

Chap. 3 SONIA-DNNChap. 3 SONIA-DNN

Page 22: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

22/56

Experimental SettingExperimental Setting

Alternative Vector

1 2 3 4 5 6 7

1 ー

2 ー ー

3 ー ー ー

4 ー ー ー ー

5 ー ー ー ー ー

6 ー ー ー ー ー ー

7 ー ー ー ー ー ー ー

21 comparison values21 comparison values

7 discarded randomly, 14 training samples7 discarded randomly, 14 training samples

Chap. 3 SONIA-DNNChap. 3 SONIA-DNN

Page 23: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

23/56

BPNN-DNN SONIA-DNN

Average Error (%)

4

2

0

Experimental ResultExperimental Result

Excellent!

Chap. 3 SONIA-DNNChap. 3 SONIA-DNN

Page 24: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

24/56

ExperimentsExperiments

AverageError (%)

Number of Samples

0

8

4

1218 15

BPNN-DNN[Chen, 2004]

SONIA-DNN[proposed]

Chap. 3 SONIA-DNNChap. 3 SONIA-DNN

JSPS Center Of Excellence Project

Wonderful!

Page 25: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

25/56

Chap. 1 IntroductionChap. 1 Introduction

ContentsContents

Chap. 2 SONIA and Food Quality PredictionChap. 2 SONIA and Food Quality Prediction

Chap. 3 SONIA-DNN for Preference ModelingChap. 3 SONIA-DNN for Preference Modeling

Chap. 4 F-SONIA for Fragrance RecognitionChap. 4 F-SONIA for Fragrance Recognition

Chap. 5 CMF-SONIA for Overlapping Pattern Class.Chap. 5 CMF-SONIA for Overlapping Pattern Class.

Chap. 6 EF-SONIA for Unknown Odor RecognitionChap. 6 EF-SONIA for Unknown Odor Recognition

Chap. 7 ConclusionsChap. 7 Conclusions

F-SONIA

Page 26: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

26/56

Perfume IndustryPerfume IndustryHuman ExpertsHuman Experts

Artificial OdorArtificial OdorDiscrimination SystemDiscrimination System

Pure PerfumePure Perfume

Two MixtureTwo Mixture

Three MixtureThree Mixture

ProblemProblemComplexityComplexity

Chap. 4 F-SONIAChap. 4 F-SONIA

Odor Discrimination SystemOdor Discrimination System

Page 27: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

27/56

Sensory System

Frequency Counter System

Neural Network

Artificial Odor Discrimination SystemArtificial Odor Discrimination SystemChap. 4 F-SONIAChap. 4 F-SONIA

Collaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

Page 28: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

28/56

SONIA : Hidden UnitSONIA : Hidden Unit

Input Vector

Unit Centroid

F-SONIA : Fuzzy Hidden UnitF-SONIA : Fuzzy Hidden Unit

Fuzzy Input Vector

Fuzzy Unit Centroid

EuclideanDistance

FuzzySimilarity

Chap. 4 F-SONIAChap. 4 F-SONIA

Collaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

FFuzzy Similarity based uzzy Similarity based SONIASONIA (1/4) (1/4) [proposed][proposed]

Page 29: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

29/56

SONIA :SONIA : Crisp Value

minimum mean maximumF-SONIA :F-SONIA :[proposed][proposed]

1

Frequency

MembershipValue

Fuzzy Triangular Number

Chap. 4 F-SONIAChap. 4 F-SONIA

Collaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

FFuzzy Similarity based uzzy Similarity based SONIASONIA (2/4) (2/4) [proposed][proposed]

Page 30: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

30/56

Similarity Value (μ)

1

MembershipValue

Frequency

Input Vector Hidden Unit Vector

Chap. 4 F-SONIAChap. 4 F-SONIA

Collaboration with University of Indonesia Under Indonesia Ministry of Sciences & Technology Project

FFuzzy Similarity based uzzy Similarity based SONIASONIA (3/4) (3/4) [proposed][proposed]

Page 31: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

31/56

SONIASONIA : :

F-SONIA :F-SONIA : [proposed][proposed]

Square Root of Quadratic Distances

Arithmetic Mean ofSimilarity Measures

Hidden Unit Input Unit

Sensor 1

Sensor i

・・・

Sensor 1

Sensor i

・・・

FFuzzy Similarity based uzzy Similarity based SONIASONIA (4/4) (4/4) [proposed][proposed]

Chap. 4 F-SONIAChap. 4 F-SONIA

Collaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

Euclidean Euclidean DistanceDistance

Fuzzy Fuzzy SimilaritySimilarity

Page 32: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

32/56

Chap. 4 F-SONIAChap. 4 F-SONIA

Citrus-Canangga-Ethanol(%)Citrus-Canangga-Ethanol(%)Collaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

F-SONIA[proposed]

Recognition (%)

100

50

0 SONIA FLVQ[Sakuraba,91]

LVQ[Kohonen,86]

BPNN[Rumelhart,86]

Page 33: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

33/56

0 100 200 300 400 500 600 700 800 900 10000

0.02

0.04

0.06

0.08

0.1

0.12

SONIA

F-SONIA

Error

Epoch

Chap. 4 F-SONIAChap. 4 F-SONIA

Error ConvergenceError Convergence

Collaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

Page 34: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

34/56

1 22 1/ 2

1 21 1 1

2 1/ 2 2

1

2 1/ 2

1

( , ) ( )

( )

( )

(

)

(

I

I

I

M M N

SONIA ai aia b i

N

bi aii

N

ai bii

g

g

g

D x x

x x

x x

2 1/ 2 2

1

( ) )IN

bi bii

g x x

SONIASONIA F-SONIAF-SONIA

1 2

22

11

1 21 1

22

21

,( , )

,

I

I

N

ai ai aiM Mi

F SONIAa b I

N

bi bi bii

I

x xD

N

x x

N

1 2 1 2( , ) ( , )SONIA F SONIAD D

Chap. 4 F-SONIAChap. 4 F-SONIADissimilarity ComparisonDissimilarity ComparisonCollaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

1 2 2

1 21 1 1

( , )HM M N

aj bja b j

D x x

Dissimilarity Definition[Hastie,01]

Page 35: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

35/56

Chap. 1 IntroductionChap. 1 Introduction

ContentsContents

Chap. 2 SONIA and Food Quality PredictionChap. 2 SONIA and Food Quality Prediction

Chap. 3 SONIA-DNN for Preference ModelingChap. 3 SONIA-DNN for Preference Modeling

Chap. 4 F-SONIA for Fragrance RecognitionChap. 4 F-SONIA for Fragrance Recognition

Chap. 5 CMF-SONIA for Overlapping Pattern Class.Chap. 5 CMF-SONIA for Overlapping Pattern Class.

Chap. 6 EF-SONIA for Unknown Odor RecognitionChap. 6 EF-SONIA for Unknown Odor Recognition

Chap. 7 ConclusionsChap. 7 Conclusions

CMF-SONIA

Page 36: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

36/56

Errors in Classification

Adaptive Clustering inspired by Adaptive Clustering inspired by B-Cell Construction of SONIA B-Cell Construction of SONIA

Class A

Class B

Chap. 5 CMF-SONIAChap. 5 CMF-SONIA

Overlapping DataOverlapping Data

Page 37: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

37/56

Chap. 5 CMF-SONIAChap. 5 CMF-SONIA

CClass lass MMajority ajority F-SONIAF-SONIA [proposed][proposed]

Class Majority for each ClusterClass Majority for each Cluster

Reduce Errors in Classification

Class A

Class B

Good Idea!

Page 38: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

38/56

Chap. 5 CMF-SONIAChap. 5 CMF-SONIA

Vowel Data Vowel Data [Lippmann,89][Lippmann,89]

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

hidhead

had

hod

heed

who’d

hawed

hudheardhood

0 750 F1(Hz)

F2(Hz)

2000

0

Page 39: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

39/56

CMF-SONIA[proposed]

Recognition (%)

80

40

0 F-SONIA BPNN[Rumelhart,86]

Excellent!

Chap. 5 CMF-SONIAChap. 5 CMF-SONIA

Recognition AccuracyRecognition Accuracy

Page 40: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

40/56

0 750 F1(Hz)

F2(Hz)

2000

0

Chap. 5 CMF-SONIAChap. 5 CMF-SONIA

Classification PlaneClassification Plane

Wow!

Page 41: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

41/56

Chap. 1 IntroductionChap. 1 Introduction

ContentsContents

Chap. 2 SONIA and Food Quality PredictionChap. 2 SONIA and Food Quality Prediction

Chap. 3 SONIA-DNN for Preference ModelingChap. 3 SONIA-DNN for Preference Modeling

Chap. 4 F-SONIA for Fragrance RecognitionChap. 4 F-SONIA for Fragrance Recognition

Chap. 5 CMF-SONIA for Overlapping Pattern Class.Chap. 5 CMF-SONIA for Overlapping Pattern Class.

Chap. 6 EF-SONIA for Unknown Odor RecognitionChap. 6 EF-SONIA for Unknown Odor Recognition

Chap. 7 ConclusionsChap. 7 Conclusions

EF-SONIA

Page 42: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

42/56

Chap. 6Chap. 6 EF-SONIAEF-SONIAUnknown Odor RecognitionUnknown Odor Recognition

Collaboration with University of IndonesiaUnder Indonesia Ministry of Sciences & Technology Project

Input

Neural NetsNeural Nets

Known Odor Unknown Odor

Page 43: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

43/56

Far with High SimilarityFar with High Similarity

High SimilarityHigh Similarity No Similarity

Arithmetic Mean

Chap. 6Chap. 6 EF-SONIAEF-SONIA

Page 44: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

44/56

Euclidean Fuzzy Similarity Euclidean Fuzzy Similarity [proposed][proposed]

NoSimilarity

Chap. 6Chap. 6 EF-SONIAEF-SONIA

Page 45: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

45/56

Similarity MeasureSimilarity Measure

Similarity Value (μ)

1

MembershipValue

Euclidean Dimension

Input Vector Hidden Unit Vector

??

??

Chap. 6Chap. 6 EF-SONIAEF-SONIA

Page 46: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

46/56

Fuzziness RegionFuzziness Region

First Dimension

Sec

ond

Dim

ensi

on

??

??

Averaging ApproachAveraging Approach

Elliptical ApproachElliptical Approach

Chap. 6Chap. 6 EF-SONIAEF-SONIA

Page 47: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

47/56

Elliptical Approach Elliptical Approach [proposed][proposed]

Θ

Brilliant Idea!

Chap. 6Chap. 6 EF-SONIAEF-SONIA

Page 48: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

48/56

Citrus-Canangga-Ethanol(%)Citrus-Canangga-Ethanol(%)

Excellent!

Method Unknown

Category Only (%)

Overall Recognition

(%)

EF-SONIA with Elliptical Approach [proposed]

96.17 98.33

EF-SONIA with Averaging Approach

89.47 96.67

Fuzzy Learning Vector Quantization (FLVQ)

[Sakuraba,91]

73.32 76.66

Learning Vector Quantization (LVQ)

[Kohonen,86]

57.63 37.91

Chap. 6Chap. 6 EF-SONIAEF-SONIA

Page 49: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

49/56

Chap. 1 IntroductionChap. 1 Introduction

ContentsContents

Chap. 2 SONIA and Food Quality PredictionChap. 2 SONIA and Food Quality Prediction

Chap. 3 SONIA-DNN for Preference ModelingChap. 3 SONIA-DNN for Preference Modeling

Chap. 4 F-SONIA for Fragrance RecognitionChap. 4 F-SONIA for Fragrance Recognition

Chap. 6 EF-SONIA for Unknown Odor RecognitionChap. 6 EF-SONIA for Unknown Odor Recognition

Chap. 7 ConclusionsChap. 7 Conclusions

SONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

Chap. 5 CMF-SONIA for Overlapping Pattern Class.Chap. 5 CMF-SONIA for Overlapping Pattern Class.

Page 50: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

50/56

Chap. 7 ConclusionsChap. 7 ConclusionsResearch ResultsResearch Results

SONIA FamilySONIA Family- Proposed Methods -- Proposed Methods -

SONIASONIA

SONIA-DNNSONIA-DNN

F-SONIAF-SONIA

CMF-SONIACMF-SONIA

- Applications -- Applications -

Food Quality PredictionFood Quality Prediction

Preference ModelingPreference Modeling

Fragrance RecognitionFragrance Recognition

Overlapping Patt. Class.Overlapping Patt. Class.

SONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

EF-SONIAEF-SONIA Unknown Odor Recog.Unknown Odor Recog.

Page 51: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

51/56

SONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

- Educational Institutes -- Educational Institutes - - Industrial Companies -- Industrial Companies -

- Governments -- Governments -

Chap. 7 ConclusionsChap. 7 ConclusionsResearch ImpactsResearch Impacts

Univ. of IndonesiaUniv. of Indonesia

Tokyo Inst. of Tech.Tokyo Inst. of Tech.

CSD Inc.CSD Inc.

IURIIURI

Japan Ministry of AgricultureJapan Ministry of Agriculture

Indonesia Ministry of Sciences & Tech.Indonesia Ministry of Sciences & Tech.

Japan Society for the Promotion of ScienceJapan Society for the Promotion of Science

Page 52: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

52/56

Related Publications Related Publications (1/5)(1/5)

Journal PapersJournal PapersM. R. Widyanto et al., “Improving Recognition and Generalization Capability of "Back-Propagation NN using a Self-Organized Network inspired by Immune Algorithm”, Applied Soft Computing Journal, Elsevier Science Pub., Vol. 6, No. 1, 2005.

[J1][J1]

M. R. Widyanto et al., “SONIA based Decision Neural Networks for Preference Assessment with Incomplete Comparisons”, International Journal of Advanced Computational Intelligence & Intelligent Informatics, Vol. 9, No. 6, 2005.

[J2][J2]

M. R. Widyanto et al., “A Fuzzy Similarity based Self-Organized Network Inspired by Immune Algorithm for Three Mixture Fragrances Recognition”, IEEE Transactions on Industrial Electronics, Vol.53, No.1, 2006 (to appear).

[J3][J3]

M. R. Widyanto et al., “Class Majority in Designing Fuzzy Local Approximation NN for Overlapping Data in Pattern Classification”, International Journal of Fuzzy Systems, Vol. 7, No. 1, 2005.

[J4][J4]

M. R. Widyanto et al., “Unknown Odor Recognition using Euclidean Fuzzy Similarity-based Self-Organized Network Inspired by Immune Algorithm”, Neural Computing and Applications, Springer-Verlag Pub., (under review).

[J5][J5]

M. R. Widyanto et al., “Local Gas Holdup Measurement using SONIA-Ultrasonic Noninvasive Method”, Sensors & Actuator – Part A: Physical, Elsevier Science Pub., Vol. 127, No.1, 2006 (to appear) .

[J6][J6]

SONIA

F-SONIASONIA-DNN CMF-SONIA

EF-SONIA

Page 53: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

53/56

Related Publications Related Publications (2/5)(2/5)

International Conference Papers (1/2)International Conference Papers (1/2)SONIA

F-SONIASONIA-DNN CMF-SONIA

EF-SONIA

M. R. Widyanto et al., “Improvement of Artificial Odor Discrimination System using Fuzzy-LVQ Neural Network”, in the proceedings of the 3rd International Conference on Computational Intelligence and Multimedia Applications, New Delhi, India, IEEE Press, pp. 474-478, 1999.

[C1][C1]

M. R. Widyanto et al., “Clustering Analysis using a Self-Organized Network Inspired by Immune Algorithm”, in the proceedings of the IASTED International Conference on Artificial and Computational Intelligence, Tokyo, Japan, ACTA Press, pp. 197-202, 2002.

[C2][C2]

M. R. Widyanto et al., “A Time-temperature-based Food Quality Prediction using a Self-Organized Network Inspired by Immune Algorithm”, in the proceedings of the 1st International Conference on Soft Computing and Intelligent Systems, Tsukuba, Japan, 2002.

[C3][C3]

M. R. Widyanto et al., “Improvement of Three mixture Fragrances Recognition using Fuzzy Similarity based Self-Organized Network Inspired by Immune Algorithm”, in the proceedings of the 4th International Symposium on Advanced Intelligent Systems, Jeju, Island, Korea, 2003.

[C4][C4]

M. R. Widyanto et al., “Class Majority in Designing a Fuzzy Local Approximation NN”, in the proceedings of the 2nd International Conference on Soft Computing and Intelligent Systems, Yokohama, Japan, 2004.

[C5][C5]

Page 54: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

54/56

Related Publications Related Publications (3/5)(3/5)

International Conference Papers (2/2)International Conference Papers (2/2)SONIA

F-SONIASONIA-DNN CMF-SONIA

EF-SONIA

M. R. Widyanto et al., “Analysis of Fuzzy Local Approximation NN on Uncertainty Decision of Frequency Measurements”, in the proceedings of the International Symposium on Computational Intelligence and Industrial Applications, Hainan, China, 2004.

[C6][C6]

M. R. Widyanto et al., “Agent-based Decision Maker Preference Modeling Using SONIA-DNN for Restaurant Work Assignment and Scheduling Problem”, in the proceedings of the International Workshop on Agent-based Approaches in Economics and Social Complex Systems, Tokyo, Japan, 2005.

[C7][C7]

M. R. Widyanto et al., “SONIA-based Decision Neural Network and Its Application to Restaurant Work Assignment”, in the proceedings of the 6th International Symposium on Advanced Intelligent Systems, Yeosu, Korea, 2005

[C8][C8]

M. R. Widyanto et al., “Unknown Odor Category Classification using EF-SONIA”, in the proceedings of the 2nd International Symposium on Computational Intelligence and Intelligence Informatics, Hammamet, Tunisia, 2005.

[C9][C9]

M. R. Widyanto et al., “SONIA-Ultrasonic Technique for Gas Holdup Measurement of a Bubble Column”, in the proceedings of the 1st Daedeok International Conference on Human-Centered Advanced Technology, Daedeok Science Town, Korea, 2005.

[C10][C10]

Page 55: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

55/56

Related Publications Related Publications (4/5)(4/5)

Domestic Conference PapersDomestic Conference Papers

M. R. Widyanto et al., “Dealing with Incomplete Comparisons using SONIA-based Decision Neural Network”, in the proceedings of the 35-th Symposium on System Engineering, Yokohama, Japan, 2005.

[D1][D1]

SONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

M. R. Widyanto et al., “Restaurant Work Assignment Modeling using SONIA-DNN”, in the proceedings of the 2nd Tokyo Tech COE RA Forum, Tokyo, Japan, 2005.

[D2][D2]

M. R. Widyanto et al., “Decision Preference Modeling using SONIA-DNN and Its Application to Work Assignment Problem, in the proceedings of the 21-th Fuzzy System Symposium, Tokyo, Japan, 2005.

[D3][D3]

Page 56: Dept. of Computational Intelligence & Systems Science Tokyo Institute of Technology

56/56

Related Publications Related Publications (5/5)(5/5)

AwardsAwards

Excellent Presentation AwardExcellent Presentation AwardThe 1st International Conference on Soft Computing & Intelligent Systems,Tsukuba, Japan, September 2002.

[A1][A1]

SONIA

F-SONIA

SONIA-DNN CMF-SONIA

EF-SONIA

Gold Prize, Best Poster Award, Master Thesis PresentationGold Prize, Best Poster Award, Master Thesis PresentationDept. of Computational Intelligence & Systems Science,Tokyo Institute of Technology, Japan, February 2003.

[A2][A2]

Outstanding Paper AwardOutstanding Paper AwardThe 6th International Conference on Advanced Intelligent Systems,Yeosu, South Korea, September 2005.

[A3][A3]