Title: Hybrid Soft Computing for Classification, Control and Prediction Applications
1Hybrid Soft Computing for Classification,
Control and Prediction Applications
- Piero P. Bonissone
- GE Global Research Center
- Bonissone_at_crd.ge.com
2Hybrid SC Applications - Outline
- Soft Computing Overview
- Introduction and SC Components
- Hybrid Soft Computing
- Modeling with FL and EA
- Hybrid SC Applications
- Prediction of Paper Web Breakage in Paper Mills
- Digital Underwriting of Insurance Products
- Conclusions
3Soft Computing
- Soft Computing (SC) the symbiotic use of many
emerging problem-solving disciplines.
- According to Prof. Zadeh
- "...in contrast to traditional hard computing,
soft computing exploits the tolerance for
imprecision, uncertainty, and partial truth to
achieve tractability, robustness, low
solution-cost, and better rapport with reality - Soft Computing Main Components
- Approximate Reasoning
- Probabilistic Reasoning, Fuzzy Logic
- Search Optimization
- Neural Networks, Evolutionary Algorithms
4Soft Computing Technologies Their Applications
at GE
- Classification
- Monitoring/Anomaly Detection
- Diagnostics
- Prognostics
- Configuration/Initialization
- Prediction
- Quality Assessment
- Equipment Life Estimation
- Scheduling
- Time/Resource Assignments
- Control
- Machine/Process Control
- Process Initialization
- Supervisory Control
- DSS/Auto-Decisioning
- Cost/Risk Analysis
- Revenue Optimization
5Problem Solving Technologies
6Hybrid SC Applications - Outline
- Soft Computing Overview
- Introduction and SC Components
- Hybrid Soft Computing
- Modeling with FL and EA
- Hybrid SC Applications
- Prediction of Paper Web Breakage in Paper Mills
- Digital Underwriting of Insurance Products
- Conclusions
7Soft Computing Probabilistic Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Bayesian Belief Nets
Dempster- Shafer
8Soft Computing Hybrid Probabilistic Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Bayesian Belief Nets
Dempster- Shafer
HYBRID PROBABILISTIC SYSTEMS
Probability of Fuzzy Events
Belief of Fuzzy Events
Fuzzy Influence Diagrams
9Soft Computing Hybrid Probabilistic Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Bayesian Belief Nets
Dempster- Shafer
HYBRID PROBABILISTIC SYSTEMS
Probability of Fuzzy Events
Belief of Fuzzy Events
Fuzzy Influence Diagrams
10Soft Computing Hybrid Probabilistic Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Bayesian Belief Nets
Dempster- Shafer
HYBRID PROBABILISTIC SYSTEMS
Probability of Fuzzy Events
Belief of Fuzzy Events
Fuzzy Influence Diagrams
11Soft Computing FL Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Fuzzy Systems
Fuzzy Logic Controllers
12Soft Computing Hybrid FL Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Fuzzy Systems
Fuzzy Logic Controllers
HYBRID FL SYSTEMS
FLC Tuned by NN (Neural Fuzzy Systems)
NN modified by FS (Fuzzy Neural Systems)
FLC Generated and Tuned by EA
13Soft Computing Hybrid FL Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Fuzzy Systems
Fuzzy Logic Controllers
HYBRID FL SYSTEMS
FLC Tuned by NN (Neural Fuzzy Systems)
NN modified by FS (Fuzzy Neural Systems)
FLC Generated and Tuned by EA
14Soft Computing Hybrid FL Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Fuzzy Systems
Fuzzy Logic Controllers
HYBRID FL SYSTEMS
FLC Tuned by NN (Neural Fuzzy Systems)
NN modified by FS (Fuzzy Neural Systems)
FLC Generated and Tuned by EA
15Soft Computing NN Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
Recurrent NN
Feedforward NN
Single/Multiple Layer Perceptron
SOM
Hopfield
ART
RBF
16Soft Computing Hybrid NN Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
Recurrent NN
Feedforward NN
Single/Multiple Layer Perceptron
SOM
Hopfield
ART
RBF
HYBRID NN SYSTEMS
NN parameters (learning rate h momentum a )
controlled by FLC
17Soft Computing Hybrid NN Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
Recurrent NN
Feedforward NN
Single/Multiple Layer Perceptron
SOM
Hopfield
ART
RBF
HYBRID NN SYSTEMS
NN topology /or
weights
generated by EAs
18Soft Computing EA Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
Genetic
Algorithms
19Soft Computing Hybrid EA Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
Evolution
Genetic
Strategies
Algorithms
Evolutionary
Genetic
Programs
Progr.
HYBRID EA SYSTEMS
EA parameters
EA parameters
(Pop size, select.)
controlled by FLC
controlled by EA
20Soft Computing Hybrid EA Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
Evolution
Genetic
Strategies
Algorithms
Evolutionary
Genetic
Programs
Progr.
HYBRID EA SYSTEMS
EA-based search
EA parameters
inter-twined with
(Pop size, select.)
hill-climbing
controlled by EA
21Soft Computing Hybrid EA Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
Evolution
Genetic
Strategies
Algorithms
Evolutionary
Genetic
Programs
Progr.
HYBRID EA SYSTEMS
EA-based search
EA parameters
inter-twined with
(Pop size, select.)
hill-climbing
controlled by EA
22Synergy in SC Reasons Approaches
- Hybrid Soft Computing
- Leverages tolerance for imprecision, uncertainty,
and incompleteness - intrinsic to the problems to
be solved - Generates tractable, low-cost, robust solutions
to such problems by integrating knowledge and
data
- Tight Hybridization
- Data-driven Tuning of Knowledge-derived Models
- Translate domain knowledge into initial structure
and parameters - Use Global or local data search to tune
parameters - Knowledge-driven Search Control
- Use Global or local data search to derive models
(Structure Parameters) - Translate domain knowledge into an algorithms
controller to improve/manage solution convergence
and quality
23Synergy in SC Reasons Approaches
- Loose Hybridization (Model Fusion)
- Combine methodologies outputs, not features
- Outputs are compared and aggregated, to increase
reliability
- Hybrid Search Methods
- Intertwining local search within global search
- Embedding knowledge in operators for global search
24Hybrid SC Applications
- Soft Computing Overview
- Introduction and SC Components
- Hybrid Soft Computing
- Modeling with FL and EA
- Hybrid SC Applications
- Prediction of Paper Web Breakage in Paper Mills
- Digital Underwriting of Insurance Products
- Conclusions
25Modeling
- Model
- Structure Parameters Search Method
-
- Classical control theory
- Structure order of the differential equations
- Parameters coefficients of differential
equation. - Search method LMSE, Pole-placement, etc.
26Soft Computing FL Systems
Approximate Reasoning
Functional Approximation/ Randomized Search
Evolutionary Algorithms
Multivalued Fuzzy Logics
Neural Networks
Probabilistic Models
Fuzzy Systems
- A Mamdani- type FLC approximates a relationship
between a state X and an output Y by using a KB
and a reasoning mechanism (generalized
modus-ponens). - The Knowledge Base (KB) is defined by
- Scaling factors (SF) ranges of values of state
and output variables - Termset (TS) membership functions of values
- Ruleset (RS) a syntactic mapping of symbols from
X to Y
Fuzzy Logic Controllers
27Example (MISO) Max-min Composition with
Centroid Defuzzification
- If X is SMALL and Y is SMALL then Z is NEG. LARGE
- If X is SMALL and Y is LARGE the Z is NEG. SMALL
- If X is LARGE and Y is SMALL the Z is POS. SMALL
- If X is LARGE and Y is LARGE then Z is POS. LARGE
Response Surface
28Modeling Using FLC (Mamdani type)
- The structure of the model is the ruleset.
- The parameters of the model are the scaling
factors and termsets. - The search method is initialized by knowledge
engineering and refined with some other external
methods (SOFC, error minimization, etc.)
29Soft Computing EA Systems
Functional Approximation/ Randomized Search
Approximate Reasoning
Multivalued Fuzzy Logics
Evolutionary Algorithms
Neural Networks
Probabilistic Models
- Most Evolutionary Algorithms (EAs) can be
described by - xt the population at time t
- under representation x
- v is the variation operator(s)
- s is the selection operator
Genetic
Algorithms
xt 1 s(v(xt))
30Evolutionary Algorithms Scalar-Valued Fitness
Function Optimization
- Example Find the maximum of the function z(x,y)
- z f(x, y) 3(1-x)2exp(-(x2) - (y1)2) -
10(x/5 - x3 - y5)exp(-x2-y2)
-1/3exp(-(x1)2 - y2).
31Modeling Using EA
- Similarly, for EA
- The structure of the model is the representation
of an individual in the population (e.g., binary
string, vector, parse tree, Finite State
Machine). - The parameters of the model are the Population
Size, Probability of Mutation, Prob. of
Recombination, Generation Gap, etc. - The search method is a global search based on
maximization of population fitness function
32Hybrid SC Applications
- Soft Computing Overview
- Introduction and SC Components
- Hybrid Soft Computing
- Modeling with FL and EA
- Hybrid SC Applications
- Prediction of Paper Web Breakage in Paper Mills
- Digital Underwriting of Insurance Products
- Conclusions
33Soft Computing Applications at GE
- GE Medical Systems
- SPT Auto Analysis for MRI (FL)
- Reverse Engineering of Picker (FL)
- FE Analysis tool (FL)
- X-Ray error Logs Analysis (CBR)
- GE Appliances
- Preferred Service Contracts (Stat.)
- Call Center Support (CBR)
- GE Capital Services
- Mortgage Collateral Evaluation
- (Fusion/FL/CBR)
- GE Aircraft Engines
- Center for Remote Diagn. (CBR)
- Customer Response Center (CBR)
- Anomaly Detection (FL/Stat.)
- IMATE - Maintenance Advisor (NN/FL)
- Resolver Drift - Sensor Fusion (FL)
Engine
- GE Financial Assurance
- GEFA LTC Preferred Customer (Stat./NN)
- GEFA Fixed Life Digital Underwriter
- (Stat, CBR, FL, GA)
- GE Transportation Systems
- Log from Transportation DB (CBR)
- Prototype Train Handling Cntrl. (FL/GA)
- Prototype Trend Analysis (Stat.)
- Embedded/Remote Diagnostics (BBN)
- GE Plastics
- Automated Color Matching (CBR)
- GE Power Gen. Systems
- Remote Anomaly Detection (Stat.)
- Embedded/Remote Diagnostics (BBN)
- Call Center Problem/Solution (CBR)
- LM Fed. Systems
- Scheduling Maintenance for
- Constellation of Satellites (GA)
- GE Industrial Systems
- Paper Web Breakage Prediction
- (NN/Stat./Induction)
- Control Mixing of Cement (FL/GA)
- LM ORSS
- Vessel Management Syst. (AI/GA)
34Enabling Soft Computing and Related Technologies
35Example Prediction of Paper Web Breakage
Developed a web breakage propensity indicator and
a time-to-break predictor for the wet-end of a
paper machine (GEIS)
Paper machine
Pulp
Paper
Press section
Dry-end
Wet-end
Water
Water
- Size of a papermaking machines 300 foot long,
3 floor high - Size of a paper web 9 to 30 feet
- Traveling speed of a web 4 to 60 MPH
- Typical monthly frequency 35-40 wet-end
breaks per machine - Peak values 15 wet-end breaks in a single
day - Average production time loss due to web break
1.6 hours per day - Average Loss of production due to web
break 512 of production - Yearly Revenue Loss Over 6 million per
machine
36System Schematics
Data Reduction
Variable Reduction
1
2
3
4
Data
Data
Variable
PCA
segmentation
scrubbing
selection
Value Transformations
Model Generation (1st Indicator)
6
5
8
7
Feature
CART
Filtering
Smoothing
Extraction
Break Indicator 1
Model Generation (2nd Indicator)
10
9
Normali-
Clustering
zation
15
13
14
11
12
Performance
Transform-
ANFIS
Trending
Shuffling
Evaluation
ation
Break Indicator 2
Time to Breaks Prediction
37Summary of Variable Selection
38Principal Components Analysis (PCA)
Data size 6,999 by 31
39The First 3 Principal Components
40Filtering First 3 Principal Components
41Feature Extraction of First Three Principal
Components
1st Principal. Comp. 2nd Principal Comp.
3rd Principal Comp.
1st PC
- Features
- 1st Derivative
- 2nd Derivative
- Difference from
- steady state
1st PC 1st Derivative.
1st PC 2nd Deriv
1st PC (Diff from SS)
3rd PC (Diff from SS)
42Classification Tree Analysis for Web Breakage
Propensity
Last 1/3 (1 hour) break indication Previous
2/3 no break indication Misclassification
10.6FP 14.8FN
43System Schematics
Data Reduction
Variable Reduction
1
2
3
4
Data
Data
Variable
PCA
segmentation
scrubbing
selection
Value Transformations
Model Generation (1st Indicator)
6
5
8
7
Feature
CART
Filtering
Smoothing
Extraction
Break Indicator 1
Model Generation (2nd Indicator)
10
9
Normali-
Clustering
zation
15
13
14
11
12
Performance
Transform-
ANFIS
Trending
Shuffling
Evaluation
ation
Break Indicator 2
Time to Breaks Prediction
44Adaptive Network-based Fuzzy Inference Systems
(ANFIS)
Inputs
IF-part
Rules
THEN-part
Output
PCA 1
x1
S
Time to break
x2
PCA 2
45Trend Analysis Time-to-break against
Predictions. A Stoplight Metaphor
46Prediction Error at time 60, i.e., E(60)Limits
for Useful Predictions and Results
47Training and Testing Process
48Synopsis Of Paper Web Prediction
- 1995-1999
- Created general process to generate Web-breakage
prediction models - Based on 1.5 year data from a paper machine in
Finland - Scope of project Wet-end breaks with unknown
causes - Project transitioned to a Finnish company 1Q
2000. - 1999-2000
- Successfully applied to wet-end web-breakage
prediction of a paper machine in Maine, USA - 2001
- Successfully applied to wet-end web-breakage
prediction of a different paper machine in Maine,
USA
49Enabling Soft Computing (and Related)
Technologies
50Example Digital Underwriting for GEFA
Developing an automated system to underwrite 50
of GEFA Fixed Life Applications by December 2,001
Future
GE Proprietary models provide uniformity
confidence in decision
51Path 1 Clean Cases (Rule Based)
EKG
APS
Summarize APS, EKG
X i
50 variables?
Medical data
Path 2
Underwriter Decision (CBE)
Policy Classification
Path 1
Application Data Received
Order APS? - Decision
Underwriter Decision (RBE)
Preferred Best Preferred Select Standard
Plus Standard Table Rating Decline
Y
X i
App
MIB data
20 variables?
X i
MVR data
Ext data
6 variables?
Paramed Lab data
Establish Confidence Start Digital Underwriting
52Fuzzy Rule Based Decision Engine (RBDE)
Preferred Best
Preferred Best
Different rule sets
give a Fuzzy Score
1
Preferred
for each rate class
Preferred
Joe Doe
Systolic Blood Pressure
New App
(for under 60 yr old)
1
Info
Systolic Blood Pressure
0.9
(for under 60 yr old)
0
Uniformity of decision
Common
140
155
160
tolerance interpretation
1
Confidence
in decision
degree of
Cholesterol
0
140
145
150
partial satisfaction of constraints
1
used to decide if human Underwriter
Cholesterol
0
needs to supervise decision
0.8
.270
.265
.270
Explanation of decision
1
traceability
to features that violated
0
D
Weight/Height
constraints (audit, compliance, etc.)
.270
275
.280
1
Immediate Applicability
encoding
D
Weight/Height
of Underwriters experience,
0
assessment, and tradeoff policies
6
3
6
DWI
NO
0
Nicotine
NO
6
8
10
Medical Conditions
NO
DWI
NO
Nicotine
NO
Medical Conditions
NO
Aggregation of Partial
Degree of
Aggregation of Partial
Degree of
Satisfaction of Constraints
Placement 0.2
Scores
determine
Aggregation of Partial
Aggregation of Partial
Degree of
best risk
Satisfaction of Constraints
Degree of
Satisfaction of Constraints
classification
Placement 0.8
53Web-based Fuzzy RBDE
54Digital Underwriting of Term Life
55Digital Underwriting of Term Life (cont.)
Met Build Req. for PREFERRED with degree
0.70 (5 lbs over)
56RBDE Parameter Tuning
Decision Engine Parameter Tuning/Optimization
Evolutionary Optimization Algorithm
Create initial population
Population of trial solutions
Proportional Selection
Intermediate population
(exploitation)
Stochastic Variation
New population
(exploration)
More superior population of trial solutions
57Decision Engine Parameter Tuning/Optimization
58Decision Engine Parameter Tuning/Optimization
Computation of Rate Class Decision Mismatch
Penalties
Decision Engine Instance
Individual k
DE Decision
Decision Comparison Correct/Incorrect
Validated Case Decision
Case
All Other Cases
59RBE Parameter Tuning via Optimization Actuarial
values (NPV) to define penalty weights
Table 1 NPV profit per policy, adjusted for
placement rate (all ages)
Table 2 Lost NPV profit per policy, adjusted
for placement rate (all ages)
60RBE Parameter Tuning via Optimization RBE
Optimization Results (Before and After)
The Optimization improved the global accuracy
from 90.1 to 93.6
61GEFA Business Impact
- Instantaneous Decisions
- Faster to Your Customer
- - Enable Creation of New Product
- - Reduce Percentage of Not-Taken policies
- Consistency with Business Rules (Lower UW
Variability) - Better Pricing, Higher Margins
- - Possibly Reduced Reinsurance Premium
- - Reduced Risk (Improved Reserve Position)
- Instant Response to Change
- Lead the Market
- - Intellectual Property Protection
- Process Control
- Six Sigma Capability
62Hybrid SC Applications
- Soft Computing Overview
- Introduction and SC Components
- Hybrid Soft Computing
- Modeling with FL and EA
- Hybrid SC Applications
- Prediction of Paper Web Breakage in Paper Mills
- Digital Underwriting of Insurance Products
- Conclusions
63Conclusions Soft Computing Applicability
- Applied SC (core related technologies) to a
broad range of applications, providing
significant benefits in both cost and growth to
many GE Businesses - Classification (Power Gen., Medical systems,
Aircraft Engine, Transportation) - Prediction (Industrial Systems)
- Scheduling (Lockheed Martin)
- Control (Industrial Systems, Trading)
- DSS /Auto-Decisioning (GE Financial Assurance,
Appliances, LM)
- Synergy In Applications of Hybrid SC
- Technology
- Combination of knowledge and data used to derive
initial structure and to refine parameters via
local or global search - Application
- Solution commonality
64Conclusions SC New Directions
- Present -gt Short Term Future
- SC technologies will widen beyond its current
constituents. - Artificial Immune Systems (for Information
Assurance) - Fractals (as building blocks in GP or for
bacteria identification) - Development of hybrid SC systems with other AI
paradigms - EA for model/software update (CBR, RBR, IR)
- Evolutionary software agents
- Push for low-cost solutions / intelligent tools
will lead to deployment of hybrid SC systems that
efficiently integrate reasoning and search
techniques.
- Medium Term Future
- SC technologies are (or will soon be) implemented
on alternative, non-standard computing mechanisms - Evolvable Hardware (Field Programmable Gate
Arrays) - Bio-inspired Systems DNA and Molecular Computing
65Conclusions SC Experiments
- Bio-inspired Systems DNA and Molecular Computing
(Examples) - Molecular Genetic Programming (Wasiewicz
Mulawka, 2001) - Representation of GP graphs by DNA molecules,
with crossover and negation operators implemented
using data flow techniques in DNA computing - DNA-based Fuzzy Systems (Deaton Garzon, 2001)
- Encoding of fuzzy membership functions in Gibbs
free energy (released upon DNA hybridization),
leading to the representation of a fuzzy rule set
(fuzzy associative memory) - Fuzzy inference performed by modified
hybridization process - DNA Neural Network Computation (Mills et al.,
2001) - DNA analog neural network in which axons and
neuron are replaced by diffusion and molecular
recognition of DNA. - DNA Evolutionary Computation (Wood et al., 2001)
- Binary-encoded Evolutionary Algorithms, with
point-wise mutation and crossover, implemented in
molecular computing, evaluating the OneMax
fitness function. - __________________________________________________
__________________________ - Watson-Crick hybridization of a pair of
complementary DNA strands makes possible a
representation of highly parallel selective
operations that is key for molecular computing
(Adleman 1994)
66Questions ?