Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Neu
ro-f
uzz
y s
yst
ems
1
Neuro-fuzzy systems: Trends and Applications
International Conference on Control, Engineering & Information Technology (CEIT’14), March 22-25, Tunisia
Dr/ Ahmad Taher AzarAssistant Professor, Faculty of Computers and
Information, Benha University, EgyptEducational Chair of IEEE RAS Egypt chapter
[email protected]: Website: http://www.bu.edu.eg/staff/ahmadazar14
Neu
ro-f
uzz
y s
yst
ems
Agenda
3
Types of models and system modeling Different Modeling Paradigms Neuro-Fuzzy systems Convergence of Technologies Advantage Of Neuro-fuzzy Modeling Types of Neuro-Fuzzy Systems Modeling With Neuro-fuzzy Systems Interpretability Versus Accuracy Of Neuro-
fuzzy Models Factors affecting the interpretability of NF
Systems Multi-adaptive Neuro Fuzzy System Design
Neu
ro-f
uzz
y s
yst
ems
4
Types of models and system modeling approaches
• Mathematical-other• Parametric-nonparametric• Continuous time- discrete time• Input-output- State space• Linear-non linear• Dynamic-Static• Time invariant-time varying• SISO -MIMO
Models
• Physical-experimental• White box-grey box-black box• Structure determination-parameter estimation• Time domain- frequency domain.
modeling–System identification
Neu
ro-f
uzz
y s
yst
ems
White Box. The model is completely constructed from a priori knowledge and
physical insight. Here, empirical data are not used during modelidentification and are only used for validation. Complete a-prioriknowledge of this kind is very rare, because usually some aspectsof the distribution of the data are unknown.
Gray Box. An incomplete model is constructed from a priori knowledge and
physical insight, then available empirical data are used to adapt themodel by finding several specific unknown parameters.
Black Box. No a priori knowledge is used to construct the model. The model is
chosen as a flexible parameterized function, which is used to fitthe data
5
Different Modeling Paradigms
Neu
ro-f
uzz
y s
yst
ems
6
Different Modeling Paradigms
Modeling Approach
Source Of Information
Method Of Acquisition Example Deficiency
Mechanistic(white-box)
Formal knowledge and
dataMathematical Differential
equations
cannot use "soft"
knowledge
Black-box Data Optimization (learning)
Regression, neural
network
cannot at all use knowledge
Fuzzy(grey box)
Various knowledge and data
Knowledge-based + learning
Rule-based model
curse” of dimensionality
Neu
ro-f
uzz
y s
yst
ems
Neuro-fuzzy systems are basically adaptive fuzzy systems developedby exploiting the similarities between fuzzy systems and certain formsof neural networks Neural Networks have their Strengths Fuzzy Logic has its Strengths
7
Neuro-Fuzzy systems
Neural Nets
KnowledgeRepresentation
Fuzzy Logic
Implicit, the systemcannot be easy interpretedor modified (-)Trains itself by learningfrom data sets (+++)
Explicit, verification andoptimization easy andefficient (+++)None, you have to defineeverything explicitly (-)Trainability
Explicit Knowledge Representation from Fuzzy Logic with TrainingAlgorithms from Neural Nets. This substantially reduces developmenttime and cost while improving the accuracy of the resulting fuzzy model.
Get “best of both worlds”:
Neu
ro-f
uzz
y s
yst
ems
8
Convergence of Technologies
Year:
1940194519501955196019651970197519801985199019952000
Computing:
Relay/Valve Based
Transistors
Small Scale Integration
Large Scale Integration
Artificial Intelligence
Neural Networks:
Neuron Model (McCulloch/Pitts)Training Rules (Hepp)
Delta Rule (Wirow/Hoff)
Multilayer Perceptron, XOR
Hopfield Model (Hopfield/Tank)Backpropagation (Rumelhart)Bidir. Assoc. Mem. (Kosko)
Fuzzy Logic:
Seminal Paper (Zadeh)
Fuzzy Control (Mamdani)
Broad Application in JapanBroad Application in EuropeBroad Application in the U.S.
Soft Computing, NeuroFuzzy
Neu
ro-f
uzz
y s
yst
ems
The advantages of neuro-fuzzy modeling are: Model identification can be performed using both empirical data
and qualitative knowledge.
The resulting models are transparent, significantly aidinghumanistic model validation and knowledge discovery.
Neuro-fuzzy systems are basically adaptive fuzzy systemsdeveloped by exploiting the similarities between fuzzy systems andcertain forms of neural networks, which fall in the class ofgeneralized local methods.
Hence, the behavior of a neuro-fuzzy system can either berepresented by a set of humanly understandable rules or by acombination of localized basis functions associated with localmodels (i.e. a generalized local method), making them an idealframework to perform nonlinear predictive modeling. model.
9
Advantage Of Neuro-fuzzy Modeling
Neu
ro-f
uzz
y s
yst
ems
There are several ways to combine neural networks and fuzzylogic.
Efforts at merging these two technologies may becharacterized by considering three main categories:
Neural Fuzzy Systems,
Fuzzy Neural Networks And
Fuzzy-neural Hybrid Systems.
10
Types of Neuro-Fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
Neural fuzzy systems are characterized by the use of neural networksto provide fuzzy systems with a kind of automatic tuning method, butwithout altering their functionality.
One example of this approach would be the use of neural networks forthe membership function elicitation and mapping between fuzzy setsthat are utilized as fuzzy rules as shown. This kind of combination ismostly used in control applications.
11
Neural Fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
In this example, the neural network simulates the processing of afuzzy system in which the neurons of the first layer are responsiblefor the fuzzification process.
12
Types of Neuro-Fuzzy Systems
The neurons of the second layerrepresent the fuzzy words usedin the fuzzy rules (third layer).
Finally, the neurons of the lastlayer are responsible for thedefuzzification process.
Neu
ro-f
uzz
y s
yst
ems
In the training process, a neural network adjusts its weights in orderto minimize the mean square error between the output of the networkand the desired output.
13
Types of Neuro-Fuzzy Systems
In this particular example, the weights ofthe neural network represent theparameters of the fuzzification function,fuzzy word membership function, fuzzyrule confidences and defuzzificationfunction respectively.
In this sense, the training of this neuralnetwork results in automatically adjustingthe parameters of a fuzzy system andfinding their optimal values.
Neu
ro-f
uzz
y s
yst
ems
Fuzzy neural Systems: The main goal of this approach is to 'fuzzify' some of the elements of
neural networks, using fuzzy logic. In this case, a crisp neuron can become fuzzy. Since fuzzy neural networks
are inherently neural networks, they are mostly used in PatternRecognition Applications.
In these fuzzy neurons, the inputs are non-fuzzy, but the weightingoperations are replaced by membership functions. The result of eachweighting operation is the membership value of the corresponding input inthe fuzzy set.
14
Types of Neuro-Fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
Hybrid neuro-fuzzy systems
In this approach, both fuzzy and neural networks techniques are usedindependently, becoming, in this sense, a hybrid system. Each one doesits own job in serving different functions in the system, incorporatingand complementing each other in order to achieve a common goal.
This kind of merging is application-oriented and suitable for bothcontrol and pattern recognition applications.
The idea of a hybrid model is the interpretation of the fuzzy rule-base in terms of a neural network. In this way the fuzzy sets can beinterpreted as weights, and the rules, input variables, and outputvariables can be represented as neurons
15
Types of Neuro-Fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
Hybrid neuro-fuzzy systems
The learning algorithm results, like in neural networks, in a change ofthe architecture, i.e. in an adaption of the weights, and/or in creatingor deleting connections. These changes can be interpreted both interms of a neural net and in terms of a fuzzy controller.
This last aspect is very important as the black box behavior of neuralnets is avoided this way. This means a successful learning procedureresults in an explicit increase of knowledge that can be represented inform of a fuzzy controller's rule base.
Hybrid neuro-fuzzy controllers are realized by approaches like ARIC,GARIC, ANFIS or the NNDFR model. These approaches consist allmore or less of special neural networks, and they are capable to learnfuzzy sets.
16
Types of Neuro-Fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
17
ANFIS (Jang 1993)
Fuzzy reasoning
A1 B1
A2 B2
w1
w2
z1 =p1*x+q1*y+r1
z2 =p2*x+q2*y+r2
z = w1+w2
w1*z1+w2*z2
x y
When Z is a first order polynomial, the resulting fuzzy inferencesystem is called a "first order Sugeno fuzzy model".
When Z is constant, the resulting model is called "zero-order Sugenofuzzy model", which can be viewed either as a special case of theMamdani inference system, in which each rule's consequent isspecified by a fuzzy singleton.
Neu
ro-f
uzz
y s
yst
ems
18
zero-order Sugeno fuzzy model
• Rule baseIf X is A1 and Y is B1 then Z = C1If X is A2 and Y is B2 then Z = C2
Neu
ro-f
uzz
y s
yst
ems
19
First-Order Sugeno FIS
• Rule baseIf X is A1 and Y is B1 then Z = p1*x + q1*y + r1
If X is A2 and Y is B2 then Z = p2*x + q2*y + r2
• Fuzzy reasoning
A1 B1
A2 B2
x=3
X
X
Y
Yy=2
w1
w2
z1 =p1*x+q1*y+r1
z =
z2 =p2*x+q2*y+r2
w1+w2
w1*z1+w2*z2
Neu
ro-f
uzz
y s
yst
ems
20
ANFIS Architecture
Layer 1: fuzzification layer Every node I in the layer 1 is an adaptivenode with a node function. Parameters in this layer: premise (orantecedent) parameters.
Layer 2: rule layer Is a fixed node labeled whose output is theproduct of all the incoming signals.
Neu
ro-f
uzz
y s
yst
ems
21
ANFIS Architecture
Layer 3: normalization layer• a fixed node labeled N.• Outputs of this layer are called normalized firing strengths.
Layer 4: defuzzification layer
• An adaptive node with a node fn O4,I = wi fi = wi (pi x + qi y + ri ) fori=1,2 where wi is a normalized firing strength from layer 3 and {pi ,qi ri } is the parameter set of this node – Consequent Parameters.
Layer 5: summation neuron• A fixed node which computes the overall output as the summation of
all incoming signals
Overall output = O5, 1 = ∑ wi fi = ∑ wi fi / ∑ wi
Neu
ro-f
uzz
y s
yst
ems
22
ANFIS Flow chart
Yes
Load Training/Testing dataGenerate initial FIS Model
• Set initial input parameters andmembership function
• Chose FIS model optimization method(hybrid method)
• Define training and testing parameters(number of training/testing epochs)
Input Training data into ANFIS system
Testing finished
Get results after training
No
Yes
Input Testing data into ANFIS system
View FIS structure,
Output surface of FIS, �
Generated rules and �
Adjusted membership functions
No
Training finished
Start
Stop
1
1
Neu
ro-f
uzz
y s
yst
ems
Remove noise/irrelevant inputs.
Remove inputs that depends on other inputs.
Make the underlying model more concise and transparent.
Reduce the time for model construction.
The selected parameters must affect the target problem, i.e.,strong relationships must exist among the parameters andtarget (or output) variables
The selected parameters must be well-populated, andcorresponding data must be as clean as possible.
23
Parameter Selection for the System
Neu
ro-f
uzz
y s
yst
ems
Whatever may be the adopted vision of fuzzy model, twodifferent phases must be carried out in fuzzy modeling,designated as structural identification. parametric identification
Structural identification consists of determining the structureof the rules, i.e. the number of rules and the number of fuzzysets used to partition each variable in the input and outputspace so as to derive linguistic labels.
Once a satisfactory structure is available, the parametricidentification must follow for the fine adjustment of theposition of all membership functions together with their shapeas the main concern
24
Modeling With Neuro-fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
Parametric Identification
Two types of parameters characterize a fuzzy model: thosedetermining the shape and distribution of the input fuzzy sets andthose describing the output fuzzy sets (or linear models).
Many neuro-fuzzy systems use direct nonlinear optimization toidentify all the parameters of a fuzzy system.
Different optimization techniques can be used to this aim. The mostwidely used is an extension of the well-known back-propagationalgorithm implemented by gradient descent. A very large number ofneuro-fuzzy systems are based on backpropagation.
25
Modeling With Neuro-fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
Hybrid training method
26
Modeling With Neuro-fuzzy Systems
A1
A2
B1
B2
S
S
/
x
y
w1
w4
w1*z1
w4*z4
Swi*zi
Swi
z
P
P
P
P
nonlinearparameters
linearparameters
Given the values of premise parameters, the overall output can beexpressed as a linear combinations of the consequent parameters.
1 21 2 1 1 2 2
1 2 1 2
1 1 1 1 1 1 2 2 2 2 2 2( ) ( ) ( ) ( ) ( ) ( )
w wf f f w f w f
w w w ww x p w y q w r w x p w y q w r
Neu
ro-f
uzz
y s
yst
ems
Hybrid training method
More specifically, in the forward pass of the hybrid learning algorithm,functional signals go forward till layer 4 and the consequentparameters are identified by the least squares estimate.
In the backward pass, the error rates propagate backward and thepremise parameters are updated by the gradient descent.
27
Modeling With Neuro-fuzzy Systems
fixed
least-squares
steepest descent
fixed
forward pass backward passMF param.(nonlinear)
Coef. param.(linear)
The consequent parameters thus identified are optimal under thecondition that the premise parameters are fixed. Accordingly the hybridapproach is much faster than the strict gradient descent and it isworthwhile to look for the possibility of decomposing the parameter set
Neu
ro-f
uzz
y s
yst
ems
Param. ID: Comparisons
Steepest descent (SD) treats all parameters as nonlinear
Hybrid learning (SD+LSE) distinguishes between linear and nonlinear
Gauss-Newton (GN) linearizes and treat all parameters as linear
Levenberg-Marquardt (LM) switches smoothly between SD and GN
28
Modeling With Neuro-fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
To speed up the process of parameter identification, many neuro-fuzzy systems adopt a multi-stage learning procedure to find andoptimize the parameters.
Typically, two stages are considered. In the first stage the input space is partitioned into regions by
unsupervised learning, and from each region the premise (andeventually the consequent) parameters of a fuzzy rule are derived.
In the second stage the consequent parameters are estimated via asupervised learning technique.
In most cases, the second stage performs also a fine adjustment ofthe premise parameters obtained in the first stage using a nonlinearoptimization technique. Most of the techniques used in theinitialization stage fall in one of the following categories:
29
Modeling With Neuro-fuzzy Systems
Neu
ro-f
uzz
y s
yst
ems
30
Grid partitioning
With this approach, the domains of the input variables are a prioripartitioned into a specified number of fuzzy sets.
The rule base is then established to cover the input space by using allpossible combinations of input fuzzy sets as multivariate fuzzy setsdescribing the rule antecedents
The consequent parameters are estimated bythe least squares method using available input-output data
Advantage: very interpretable fuzzy sets canbe generated.
Drawback: the number of multivariate fuzzysets, and hence the number of rules, is anexponential function of the number of inputs.
This curse of dimensionality restricts the use of fuzzy models based ongrid partitioning to low dimensional problems.
Neu
ro-f
uzz
y s
yst
ems
31
Cluster-oriented methods
Cluster-oriented methods try to group the training data into clusters anduse them to define multivariate fuzzy sets describing the premise part offuzzy rules
Popular clustering methods adopted to find the centers of multivariatemembership functions are the k-means clustering, the self-organizingfeature maps (SOM), Fuzzy clustering.
While clustering-based methods produce verya flexible partitioning with respect to thegrid-based approaches, the lattice partition ofthe input space is ignored and this usuallyresults in rule bases that cannot beinterpreted very well.
As shown, the projection of clusters toobtain fuzzy rules typically results inoverlapping nonsensical fuzzy sets that arehard to interpret linguistically
Neu
ro-f
uzz
y s
yst
ems
32
Cluster-oriented methods
In the case of fuzzy clustering, the projection causes a loss ofinformation because the Cartesian product of the induced membershipfunctions does not reproduce the fuzzy cluster exactly
Another consequence of cluster projection is that for each variablethe fuzzy sets are induced individually for each rule and for eachfeature there will be as many different fuzzy sets as the number ofclusters.
Some of these fuzzy sets may be similar, yet they are usually notidentical. For a good interpretation it is necessary to have a fuzzypartition of few fuzzy sets where each clearly represents a linguisticconcept.
To eliminate this redundancy, similarity measures can be used in orderto assess the degree of overlapping between adjacent fuzzy sets andmerge fuzzy sets that are too similar.
Neu
ro-f
uzz
y s
yst
ems
33
Structural Identification
Before fuzzy rule parameters can be optimized, the structure of thefuzzy rule base must be defined. This involves determining the numberof rules and the granularity of the data space, i.e. the number of fuzzysets used to partition each variable.
In fuzzy rule-based systems, as in any other modeling technique, thereis a tradeoff between accuracy and complexity.
The more rules, the finer the approximation of the nonlinear mappingcan be obtained by the fuzzy system, but also more parameters haveto be estimated, thus the cost and complexity increase
A possible approach to structure identification is to perform astepwise search through the fuzzy model space. Once again, thesesearch strategies fall into one of two general categories: forwardselection and backward elimination.
Neu
ro-f
uzz
y s
yst
ems
34
Structural Identification
Forward selection. Starting from a very simple rule base, new fuzzyrules are dynamically added or the density of fuzzy sets isincrementally increased.
Backward elimination. An initial fuzzy rule base, constructed from apriori knowledge or by learning from data, is reduced, until a minimumof the error function is found. The structure of the fuzzy rules canalso be optimized by GA's so that a compact fuzzy rule base can beobtained
The learning algorithm is an example of structure adaptation in neuro-fuzzy systems. Rules are dynamically recruited or deleted according totheir significance to system performance, so that a parsimoniousstructure with high performance is achieved.
Neu
ro-f
uzz
y s
yst
ems
35
Interpretability Versus Accuracy Of Neuro-fuzzy Models
The twofold face of fuzzy systems leads to a trade-off betweenreadability and accuracy
Interpretability Accuracy
No. of parameters Few Parameters More Parameters
No. of fuzzy rules Few Rules More Rules
Type of Fuzzy logic Model Mamdani Models TSK models
To keep the model simple, the prediction is usually less accurate. Insolving this trade-off the interpretability (meaning also simplicity) offuzzy systems must be considered the major advantage and hence itshould be pursuit more than accuracy.
In fact fuzzy systems are not better function approximators orclassifiers than other approaches. If we are interested in a veryprecise prediction, then we are usually not so much interested in theinterpretability of the solution
Neu
ro-f
uzz
y s
yst
ems
36
Factors affecting the interpretability of NF Systems
Choice of fuzzy model type: Mamdani-type (or singleton) fuzzy systems should be preferred to
TS fuzzy systems, because the rule, consequents consist ofinterpretable fuzzy sets.
Number of fuzzy rules: a fuzzy system with a large rule base is less interpretable than a
fuzzy system that needs only few rules.
Number of input variables: each rule should use as few variables as possible to be more
comprehensible.
Number of fuzzy sets per variable: only a moderate number of fuzzysets should be used to partition a variable. A coarse granularityincreases the readability of the fuzzy model, hence too many linguisticlabels for each variable is preferentially to avoid.
Neu
ro-f
uzz
y s
yst
ems
37
Factors affecting the interpretability of NF Systems
Characteristics of fuzzy rules: the fuzzy rules must be complete, i.e. for any input, the rule-based
system can generate an answer. Also the rule base must be consistent, i.e. there must be no
contradictory rules that have identical antecedents but differentconsequents. Only partial inconsistency is acceptable.
Also, any form of redundancy should be avoided, e.g. there must beno rule whose antecedent is a subset of the antecedent of anotherrule, and no rule may appear more than once in the rule base.
Characteristics of fuzzy sets: fuzzy sets should be "interpretable" tothe user of the fuzzy system. This means that membership functionsshould be normal, convex and they should guarantee a completecoverage of the corresponding input domain (coverage), so that a usershould be able to label each fuzzy set by a linguistic term. Also toomuch overlapping between the membership functions should beprevented, so as to have distinguishable fuzzy sets.
Neu
ro-f
uzz
y s
yst
ems
38
Multi-adaptive Neuro Fuzzy System Design
Ensemble-Based Approach
Combination Module
NF Network 1
NF Network 2
NF Network N
Training Set
Neu
ro-f
uzz
y s
yst
ems
39
Multi-adaptive Neuro Fuzzy System Design
Modular-Based Approach
Sub-Task N
Subset 1
Subset 2
Subset N
Sub-Task 1 Sub-Task 2
NF Network 1
NF Network 2
NF Network N
Decomposition of the training set into N different groups
Decomposition of the task into N different sub-tasks
TrainingSet
CombinationModule
Neu
ro-f
uzz
y s
yst
ems
40
Multi-adaptive Neuro Fuzzy System Design
Hybrid-Based neuro-fuzzy combination approach
Combination Module
NF Network 2
Training Set
Sub-Task 1
ModularModule
Sub-Task 2
NF Network 2
NF Network 2
Sub-Task N
Ensemble Module
Neu
ro-f
uzz
y s
yst
ems
41
Conclusion
Neuro-fuzzy modeling approaches combine the benefits of twopowerful paradigms into a single capsule and provide a powerfulframework to extract fuzzy (linguistic) rules from numerical data.
The aim of using a neuro-fuzzy network is to find, through learningfrom data, a fuzzy model that represents the process underlying thedata.
Contributing factors to successful applications of neuro-fuzzy andsoft computing: Sensor technologies
Cheap fast microprocessors
Modern fast computers
Neu
ro-f
uzz
y s
yst
ems
42
Multi-adaptive Neuro Fuzzy System Design
References: “Neuro-Fuzzy and Soft Computing”, J.-S. R. Jang, C.-T. Sun and
E. Mizutani, Prentice Hall, 1996 “Neuro-Fuzzy Modeling and Control”, J.-S. R. Jang and C.-T. Sun,
the Proceedings of IEEE, March 1995. “ANFIS: Adaptive-Network-based Fuzzy Inference Systems,”,
J.-S. R. Jang, IEEE Trans. on Systems, Man, and Cybernetics, May 1993.
Internet resources: This set of slides is available at
http://www.cs.nthu.edu.tw/~jang/publication.htm WWW resouces about neuro-fuzzy and soft-computing
http://www.cs.nthu.edu.tw/~jang/nfsc.htm
Neu
ro-f
uzz
y s
yst
ems
43