Fuzzy Modelling - COST Action IC0702
Fuzzy Modelling - COST Action IC0702
Fuzzy Modelling - COST Action IC0702
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Fuzzy</strong> <strong>Modelling</strong>:<br />
Fundamentals, Design, and<br />
Challenges<br />
IFSA 2009<br />
Lisbon, July 20, 2009
Roadmap<br />
Overview<br />
Motivation<br />
Fundamental quests<br />
Type-2 fuzzy<br />
models and<br />
interpretability<br />
General architecture<br />
and functional modules<br />
General architecture<br />
Functional modules<br />
<strong>Fuzzy</strong> models and fuzzy modeling<br />
Graph-oriented<br />
fuzzy models<br />
Verification<br />
& Validation<br />
(V & V)<br />
Multimodal and<br />
collaborative fuzzy<br />
models<br />
Direction-based<br />
fuzzy models<br />
Design of<br />
Information granules<br />
Incremental<br />
fuzzy models<br />
Linguistic<br />
models
<strong>Fuzzy</strong> models – historical perspective<br />
<strong>Fuzzy</strong> models<br />
Neurofuzzy models<br />
Hybrid fuzzy models<br />
FUZZY SETS<br />
COMPUTATIONAL INTELLIGENCE
<strong>Fuzzy</strong> Models: some<br />
statistics<br />
query : fuzzy model and neurofuzzy<br />
35000<br />
30000<br />
25000<br />
20000<br />
15000<br />
10000<br />
5000<br />
0<br />
1 3 5 7 9 11 13 15 17 19 21 23 25 27<br />
1970 1980<br />
2009<br />
Google Scholar, April 2009
<strong>Fuzzy</strong> modeling: an overview<br />
Plethora of methodologies and architectures of fuzzy models<br />
Hybrid design strategies (fuzzy, neurofuzzy, evolutionary<br />
techniques)<br />
Predominantly numeric nature of results (fuzzy models with numeric<br />
decoding modules)<br />
Multiobjective nature of fuzzy models with several conflicting criteria<br />
•Accuracy<br />
•Interpretability<br />
•stability
<strong>Fuzzy</strong> modeling:<br />
Fundamental quests<br />
Dominant role of designer (user) in system modeling<br />
Sound design practices which help the designer assume<br />
active role throughout whole development process<br />
Models easily adjustable to current requirements imposed by<br />
the problem; effective realization of tradeoffs between<br />
accuracy and interpretability<br />
Successive refinements of models<br />
Controlled design effort (avoidance of excessively long learning, …)
<strong>Fuzzy</strong> models with<br />
Information granules: a retrospective<br />
Information granules are formed in multivariable input space.<br />
With each of them comes some local model associating<br />
inputs with the output<br />
curse of dimensionality<br />
Examples - Rule-based systems<br />
If cond 1<br />
is A 1<br />
and cond 2<br />
is A 2<br />
… then conclusion is B<br />
If cond 1<br />
is A 1<br />
and cond 2<br />
is A 2<br />
… then conclusion is f(x, a)<br />
If (cond 1<br />
, cond 2<br />
…cond 3<br />
) is R then conclusion is B<br />
Relational constraint
Example<br />
If (cond 1<br />
, cond 2<br />
…cond n<br />
) is R then conclusion is B<br />
If (cond 1<br />
, cond 2<br />
…cond n<br />
) is R then conclusion is f(x,a)<br />
Information granules<br />
aggregation
<strong>Fuzzy</strong> sets and interfaces<br />
<strong>Fuzzy</strong> sets (and sets) do not exist in real-world<br />
To interact with the world one has to construct interfaces<br />
(encoders and decoders)<br />
Encoder<br />
<strong>Fuzzy</strong> set-based<br />
processing<br />
Decoder<br />
Interfacing
Digital processing: an analogy<br />
D/A<br />
Digital<br />
Processing<br />
A/D
Functional modules of interfaces<br />
Encoders The objective is to translate input data into some internal format<br />
acceptable for processing at level of fuzzy sets<br />
Decoders The objective is to convert the results of processing of fuzzy sets into<br />
some format acceptable by the external world (typically in the form of some<br />
numeric quantities)<br />
For encoding and decoding we engage a collection of fuzzy sets – information<br />
granules
Encoding<br />
Given is a collection of fuzzy sets A 1<br />
, A 2<br />
, …, A c<br />
; express<br />
some numeric input x in R in terms of these fuzzy sets<br />
x [ A 1<br />
(x) A 2<br />
(x)… A c<br />
(x)]<br />
Nonlinear mapping from R to c-dimensional unit hypercube
Decoding<br />
(a) decoding completed on a basis of a single fuzzy set<br />
(b) Decoding realized on a basis of a certain finite family of fuzzy sets and<br />
levels of their activation.
Decoding process: a single fuzzy set<br />
Single fuzzy set B develop a single numeric representative
Single fuzzy set decoding: centre of gravity<br />
Solution to the following optimization problem<br />
V<br />
=<br />
∫<br />
X<br />
B(x)[x<br />
−<br />
xˆ] 2 dx<br />
dV = 0<br />
2<br />
dxˆ<br />
∫ B(x)[x − xˆ]dx =<br />
X<br />
0
Single fuzzy set decoding: augmented strategies<br />
Augmented centre of gravity<br />
xˆ<br />
=<br />
∫<br />
x∈X:B(x)<br />
≥β<br />
∫<br />
x∈X:B(x)<br />
≥β<br />
B(x)xdx<br />
B(x)dx<br />
xˆ<br />
=<br />
∫<br />
x∈X:B(x)<br />
≥β<br />
∫<br />
B<br />
γ<br />
B<br />
γ<br />
x∈X:B(x)<br />
≥β<br />
(x)xdx<br />
(x)dx
Decoding: a collection of fuzzy sets<br />
x<br />
xˆ<br />
ENCODER<br />
DECODER<br />
Numeric<br />
Input<br />
(multidimensional)<br />
Granular<br />
representation<br />
Numeric<br />
Output<br />
(multidimensional)<br />
•One-dimensional case<br />
•Multivariable case- to be studied later
Decoding: one-dimensional (scalar) case<br />
Codeboook – a finite family of fuzzy sets {A 1<br />
, A 2<br />
, …, A c<br />
}<br />
A 1 A 2 A i A i+1<br />
1/2<br />
v i v i+1<br />
x
Design of Information Granules<br />
Development of multivariable fuzzy sets (information granules)<br />
R 1<br />
, R 2<br />
, …, R c<br />
<strong>Fuzzy</strong> clustering as a constructive vehicle of forming<br />
information granules
<strong>Fuzzy</strong> Clustering: <strong>Fuzzy</strong> C-Means (FCM)<br />
Given data x 1<br />
, x 2<br />
, …, x N<br />
, determine its structure by<br />
forming a collection of information granules – fuzzy sets<br />
Objective function<br />
Q<br />
=<br />
c<br />
∑<br />
i=<br />
1<br />
N<br />
∑<br />
k=<br />
1<br />
u<br />
m<br />
ik<br />
||<br />
x<br />
k<br />
−<br />
v<br />
i<br />
||<br />
2
FCM – representation fundamentals<br />
c<br />
∑<br />
i=1<br />
0 <<br />
u ik<br />
=1, k =1,2,..., N<br />
N<br />
∑ uik < N, i =<br />
k=<br />
1<br />
1,2,..., c
FCM – optimization<br />
Q<br />
=<br />
c<br />
∑<br />
i=<br />
1<br />
N<br />
∑<br />
k=<br />
1<br />
u<br />
m<br />
ik<br />
||<br />
x<br />
k<br />
−<br />
v<br />
i<br />
||<br />
2<br />
Minimize<br />
subject to<br />
(a) prototypes<br />
(b) partition matrix
Optimization - details<br />
Partition matrix – the use of Lagrange multipliers<br />
V =<br />
c<br />
∑<br />
m<br />
u ik<br />
d ik<br />
i=1<br />
c<br />
∑<br />
i=1<br />
2 + λ( u ik<br />
−1)<br />
d ik = ||x k -v i || 2<br />
λ –Lagrange multiplier<br />
∂V<br />
= 0 ∂V<br />
∂u st<br />
∂λ = 0
Optimization – partition matrix (1)<br />
∑<br />
∑<br />
=<br />
=<br />
−<br />
+<br />
=<br />
c<br />
1<br />
i<br />
ik<br />
2<br />
ik<br />
c<br />
1<br />
i<br />
m<br />
ik 1)<br />
u<br />
λ(<br />
d<br />
u<br />
V<br />
0<br />
λ<br />
V<br />
0<br />
u<br />
V<br />
st<br />
=<br />
∂<br />
∂<br />
=<br />
∂<br />
∂<br />
λ<br />
d<br />
mu<br />
u<br />
V<br />
2 st<br />
1<br />
m<br />
st<br />
st<br />
+<br />
=<br />
∂<br />
∂<br />
−<br />
d<br />
m<br />
λ<br />
u 1<br />
m-<br />
2<br />
st<br />
m-1<br />
1<br />
st<br />
⎟ ⎠ ⎞<br />
⎜<br />
⎝<br />
⎛<br />
−<br />
= ∑ =<br />
⎟<br />
⎠<br />
⎞<br />
⎜<br />
⎝<br />
⎛<br />
−<br />
=<br />
−<br />
− c 1<br />
j<br />
1<br />
m<br />
2<br />
jt<br />
1<br />
m<br />
1<br />
1<br />
d<br />
m<br />
λ<br />
∑<br />
=<br />
⎟<br />
⎠<br />
⎞<br />
⎜<br />
⎝<br />
⎛<br />
−<br />
=<br />
−<br />
−<br />
c<br />
1<br />
j<br />
1<br />
m<br />
2<br />
jt<br />
1<br />
m<br />
1<br />
d<br />
1<br />
m<br />
λ<br />
∑<br />
=<br />
−<br />
⎟<br />
⎟<br />
⎠<br />
⎞<br />
⎜<br />
⎜<br />
⎝<br />
⎛<br />
=<br />
c<br />
1<br />
j<br />
1<br />
m<br />
1<br />
2<br />
jt<br />
2<br />
st<br />
st<br />
d<br />
d<br />
1<br />
u
Optimization- prototypes (2)<br />
Q<br />
=<br />
c<br />
∑<br />
i=<br />
1<br />
N<br />
∑<br />
k=<br />
1<br />
u<br />
m<br />
ik<br />
n<br />
∑ (x<br />
j=<br />
1<br />
kj<br />
−<br />
v<br />
ij<br />
)<br />
2<br />
Euclidean distance<br />
Gradient of Q with respect to v s ∑ u (x − v ) =<br />
N<br />
k=<br />
1<br />
m<br />
ik<br />
kt<br />
st<br />
0<br />
v<br />
st<br />
=<br />
N<br />
∑<br />
u<br />
k=<br />
1<br />
N<br />
∑<br />
k=<br />
1<br />
m<br />
ik<br />
u<br />
x<br />
m<br />
ik<br />
kt
<strong>Fuzzy</strong> C-Means (FCM): An overview<br />
procedure FCM-CLUSTERING (x) returns prototypes and partition matrix<br />
input : data x = {x 1, x 2, ..., x k}<br />
local: fuzzification parameter: m<br />
threshold: ε<br />
norm: ||.||<br />
INITIALIZE-PARTITION-MATRIX<br />
t ← 0<br />
repeat<br />
for i=1:c do<br />
N<br />
m<br />
∑ u<br />
ik<br />
(t) xk<br />
v ←<br />
k = 1<br />
i (t)<br />
compute prototypes<br />
N<br />
m<br />
∑ uik<br />
(t)<br />
k = 1<br />
for i = 1:c do<br />
for k = 1:N do<br />
update partition matrix<br />
1<br />
uik<br />
(t + 1) =<br />
2/(m−1)<br />
c ⎛<br />
⎞<br />
∑ ⎜<br />
|| xk<br />
− vi<br />
(t) ||<br />
⎟<br />
j=<br />
1<br />
⎜<br />
⎟<br />
⎝<br />
|| xk<br />
− v j(t)<br />
||<br />
⎠<br />
update partition matrix<br />
t ← t + 1<br />
until ||U(t+1)-U(t)|| ≤ ε<br />
return U, V
Geometry of information granules<br />
n=1<br />
1<br />
1<br />
1<br />
1<br />
1<br />
1<br />
A( x,<br />
1.2)<br />
A( x,<br />
2)<br />
A( x,<br />
3.5)<br />
B( x,<br />
1.2)<br />
0.5<br />
B( x,<br />
2)<br />
0.5<br />
B( x,<br />
3.5)<br />
0.5<br />
C( x,<br />
1.2)<br />
C( x,<br />
2)<br />
C( x,<br />
3.5)<br />
0<br />
0<br />
1 2 3<br />
0.5 x<br />
3.5<br />
2.265×<br />
10 − 7<br />
0<br />
1 2 3<br />
0.5 x<br />
3.5<br />
2.18×<br />
10 − 3<br />
0<br />
1 2 3<br />
0.5 x<br />
3.5<br />
m =1.2 m =2.0 m =3.5
<strong>Fuzzy</strong> Clustering: choosing granularity of<br />
information granules<br />
Cluster validity measures…<br />
Reconstruction criterion<br />
Given original numeric datum x k<br />
, express it in terms of<br />
clusters (information granules) and re-construct it.<br />
The reconstruction error is a measure expressing<br />
differences between original datum and its reconstruction<br />
V<br />
N<br />
= ∑||<br />
k=<br />
1<br />
x −<br />
ˆ<br />
2<br />
k<br />
xk<br />
||
Multivariable encoding and decoding: a global<br />
view<br />
Encoding<br />
i 0<br />
Decoding<br />
{v 1 , v 2 , …, v c }<br />
VQ<br />
use of sets –<br />
Vector Quantization<br />
(VQ)<br />
Prototypes<br />
v 1 , v 2 , …, v c<br />
Prototypes<br />
v 1 , v 2 , …, v c<br />
Encoding<br />
u 1 , u 2 , …,u c-1<br />
Decoding<br />
FVQ<br />
use of fuzzy sets –<br />
<strong>Fuzzy</strong> Vector Quantization<br />
(FVQ)<br />
Prototypes<br />
v 1 , v 2 , …, v c<br />
Prototypes<br />
v 1 , v 2 , …, v c
<strong>Fuzzy</strong> Vector Quantization<br />
The codebook formed through fuzzy clustering (FCM) producing<br />
A finite collection of prototypes v 1<br />
, v 2<br />
, …, v c<br />
.<br />
Given any new input x we realize its encoding and decoding<br />
Let us recall:<br />
Encoding – representation of x in terms of the prototypes<br />
Decoding – development of external representation of the result of<br />
processing realized at the level of information granules
u<br />
i<br />
( x)<br />
∈[0,1],<br />
∑ u<br />
i<br />
(x) = 1<br />
c<br />
i=<br />
1<br />
<strong>Fuzzy</strong> Vector Quantization:<br />
Encoding<br />
The optimization problem<br />
c<br />
∑ u<br />
i=<br />
1<br />
m<br />
2<br />
i<br />
|| x − vi<br />
||<br />
Minimize w.r.t. u i<br />
subject to<br />
u<br />
i<br />
( x)<br />
∈[0,1],<br />
∑ u<br />
i<br />
(x) = 1<br />
c<br />
i=<br />
1<br />
u<br />
i<br />
( x)<br />
=<br />
⎛<br />
∑⎜<br />
||<br />
⎝ ||<br />
1<br />
x − v<br />
x − v<br />
i<br />
j<br />
|| ⎞<br />
⎟<br />
||<br />
⎠<br />
2<br />
m−1
u<br />
i<br />
( x)<br />
∈[0,1],<br />
∑ u<br />
i<br />
(x) = 1<br />
c<br />
i=<br />
1<br />
<strong>Fuzzy</strong> Vector Quantization:<br />
Decoding<br />
Reconstruct original mutidimensional input x<br />
c<br />
m<br />
2<br />
2<br />
(ˆ) x = u ˆ<br />
i<br />
|| x v<br />
i<br />
||<br />
i=<br />
1<br />
Q ∑ −<br />
minimize<br />
xˆ<br />
=<br />
c<br />
∑ u<br />
i=<br />
1<br />
c<br />
∑ u<br />
i=<br />
1<br />
m<br />
i<br />
v<br />
m<br />
i<br />
i
u<br />
i<br />
( x)<br />
∈[0,1],<br />
∑ u<br />
i<br />
(x) = 1<br />
c<br />
i=<br />
1<br />
<strong>Fuzzy</strong> Vector Quantization:<br />
Decoding error
Example – computing<br />
Input-output relationship<br />
z 1<br />
z i<br />
z c<br />
y<br />
=<br />
c<br />
∑<br />
i=<br />
1<br />
z<br />
i<br />
u<br />
i<br />
( x)<br />
u ( x)<br />
=<br />
i<br />
c<br />
∑<br />
j=<br />
1<br />
⎛<br />
⎜<br />
|| x<br />
⎝ || x<br />
1<br />
−<br />
−<br />
v<br />
v<br />
i<br />
j<br />
|| ⎞<br />
⎟<br />
||<br />
⎠<br />
2<br />
m−1
Examples (1)<br />
y<br />
5<br />
m=1.2<br />
m=4.0<br />
v 1<br />
= -1, v 2<br />
= 2.5 v 3<br />
= 6.1; z 1<br />
= 6, z 2<br />
= -4, z 3<br />
= 2<br />
0<br />
m=2.0<br />
5<br />
− 5<br />
4 2 0 2 4<br />
x<br />
x<br />
5<br />
y<br />
5<br />
y<br />
5<br />
Change of prototypes<br />
in input space<br />
0<br />
0<br />
5<br />
0 2 4 6<br />
− 0<br />
x<br />
x<br />
7<br />
5<br />
0 2 4 6<br />
x
<strong>Fuzzy</strong> Models: A General Taxonomy<br />
Source of data/knowledge<br />
single source<br />
Multiple sources<br />
Direction –based<br />
architectures<br />
Graph–oriented<br />
architectures
The architectural blueprint<br />
of<br />
fuzzy models
Preamble<br />
As modeling is realized at higher, more abstract level, fuzzy<br />
models give rise to a general architecture in which we<br />
highlight three main functional modules, that is<br />
– input interface<br />
– processing module<br />
– output interface
A general architecture<br />
<strong>Fuzzy</strong> model<br />
Domain<br />
knowledge<br />
Processing<br />
Interface<br />
Interface<br />
Data<br />
Decision, control signal,<br />
class assignment…
Main categories of fuzzy<br />
models: An overview
Main categories of models:<br />
An overview<br />
Diversified landscape of fuzzy models - selected categories:<br />
– tabular fuzzy models<br />
– rule-based fuzzy models<br />
– fuzzy relational models including associative memories<br />
– fuzzy decision trees<br />
– fuzzy neural networks<br />
– fuzzy cognitive maps<br />
– ….
Some design considerations<br />
• Expressive power<br />
• Processing capabilities<br />
• Design schemes and ensuing optimization<br />
• Interpretability<br />
• Ability to deal with heterogeneous data<br />
• ….
Tabular fuzzy models<br />
• Table of relationships between the variables of the system<br />
granulated by some fuzzy sets.<br />
• Easy to build and interpret<br />
• Limited processing capabilities (not included as a part of the model)<br />
B 1 B 2 B 3 B 4 B 5<br />
A 1<br />
A 2<br />
C 3<br />
A 3<br />
C 1
Rule-based fuzzy models<br />
• Highly modular and easily expandable fuzzy models<br />
• Composed of a family of conditional (If – then) statements (rules)<br />
• <strong>Fuzzy</strong> sets occur in their conditions and conclusions<br />
• Standard format<br />
If condition 1<br />
is A and condition 2<br />
is B and … and<br />
condition n<br />
is W<br />
then conclusion is Z<br />
• Conditions ≡ rule antecedent<br />
• Conclusions ≡ rule consequent
Rule-based fuzzy models:<br />
granularity and quality of rules<br />
Low High<br />
granularity of conclusion<br />
general condition<br />
(highly applicable rule)<br />
and very specific<br />
conclusion. High quality<br />
rule<br />
high generality of the<br />
rule, low specificity of the<br />
conclusion, average<br />
quality of the rule<br />
Low High<br />
granularity of condition<br />
condition and conclusion<br />
highly specific; lack of<br />
generalization; very<br />
limited relevance of the<br />
rule<br />
limited generality<br />
(specific condition) and<br />
lack of specificity of<br />
conclusion; low quality<br />
rule
<strong>Fuzzy</strong> rule-based model: design<br />
If x is B i then y = f i (x, p i ), i=1, 2, …,c<br />
Determination of condition parts (information granules) of the rules:<br />
<strong>Fuzzy</strong> clustering (e.g., <strong>Fuzzy</strong> C-Means, FCM) commonly<br />
Used to build information granules B i<br />
Determination of conclusion parts of the rules:<br />
Estimation of parameters (p i ) –optimization problem;<br />
global minimum of the problem could be achieved
Rule-based fuzzy models<br />
fuzzy rule-based models with linear local models<br />
-if x is B i then y = a iT x<br />
Output as an aggregation of local models and activation<br />
levels<br />
ŷ<br />
c<br />
k = ∑ ui<br />
( xk<br />
)<br />
i=<br />
1<br />
a<br />
T<br />
i<br />
x<br />
k
Rule-based fuzzy models:<br />
development (1)<br />
2<br />
N<br />
1<br />
k<br />
k<br />
k )<br />
ŷ<br />
Q ∑ (y<br />
=<br />
−<br />
=<br />
k<br />
T<br />
i<br />
k<br />
c<br />
1<br />
i<br />
i<br />
k )<br />
(<br />
u<br />
ŷ<br />
x<br />
a<br />
x<br />
∑<br />
=<br />
=<br />
∑<br />
=<br />
−<br />
⎟<br />
⎟<br />
⎠<br />
⎞<br />
⎜<br />
⎜<br />
⎝<br />
⎛<br />
−<br />
−<br />
=<br />
c<br />
1<br />
j<br />
1<br />
m<br />
2<br />
j<br />
k<br />
i<br />
k<br />
ik<br />
||<br />
||<br />
||<br />
||<br />
1<br />
u<br />
v<br />
x<br />
v<br />
x<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
ck<br />
2k<br />
1k<br />
T<br />
c<br />
T<br />
2<br />
T<br />
1<br />
k ]<br />
...<br />
[<br />
ŷ<br />
z<br />
z<br />
z<br />
a<br />
a<br />
a<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
c<br />
2<br />
1<br />
a<br />
a<br />
a<br />
a<br />
k<br />
k<br />
i<br />
ik )<br />
(<br />
u<br />
x<br />
x<br />
z =
Rule-based fuzzy models:<br />
development (2)<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
N<br />
2<br />
1<br />
y<br />
y<br />
y<br />
y<br />
)<br />
(Z<br />
)<br />
(Z<br />
Q<br />
T<br />
y<br />
a<br />
y<br />
a<br />
−<br />
−<br />
=<br />
y<br />
a<br />
T<br />
1<br />
T<br />
opt<br />
Z<br />
Z)<br />
(Z<br />
−<br />
=<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎦<br />
⎤<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎣<br />
⎡<br />
=<br />
cN<br />
c2<br />
c1<br />
1N<br />
12<br />
11<br />
Z<br />
z<br />
z<br />
z<br />
z<br />
z<br />
z
<strong>Fuzzy</strong> relational structures:<br />
A general taxonomy<br />
t-conorms<br />
nullnorms<br />
infimum (min)<br />
uninorms<br />
min-uninorm composition<br />
t-norms<br />
inf-s composition<br />
ordinal sum<br />
max-min composition<br />
supremum (max)<br />
sup-min composition<br />
implications<br />
sup-t composition
<strong>Fuzzy</strong> decision trees<br />
• Generalization of decision trees<br />
A={a 1 , a 2 , a 3 }<br />
∈<br />
B={b 1 , b 2 } C={c 1 , c 2 , c 3 , c 4 }<br />
≥<br />
a 3 , c 1<br />
• Traversal of tree depending on the values of the attributes:<br />
only a single path traversed and a single terminal node<br />
reached
<strong>Fuzzy</strong> decision trees<br />
• Traversal of a number of paths leading to a number of<br />
terminal nodes (reachability levels)<br />
A = {A 1 , A 2 , A 3 }<br />
B = {B 1 , B 2 } C = {C 1 , C 2 , C 3 , C 4 }<br />
µ 1 µ 2 µ 3 µ 4 µ 5 µ 6 reachability
<strong>Fuzzy</strong> decision trees<br />
• Traversal of a number of paths leading to a number of<br />
terminal nodes (reachability levels)<br />
x<br />
A = {A 1 , A 2 , A 3 }<br />
C = {C 1 , C 2 , C 3 , C 4 }<br />
y<br />
µ = A 1 (x) t C 2 (y) reachability
<strong>Fuzzy</strong> neural networks<br />
• Architectures in which we combine adaptive properties of<br />
neural networks with interpretability (transparency) of<br />
fuzzy sets<br />
• A suite of fuzzy logic neurons:<br />
– aggregative neurons (and, or neurons)<br />
– referential neurons (dominance, equality, inclusion…)<br />
• Learning mechanisms could be applied to adjustment<br />
of connections of neurons<br />
• Each neuron comes with a well-defined semantics; the<br />
network could be easily interpreted once the training has<br />
been completed
<strong>Fuzzy</strong> neural networks:<br />
Examples of architectures<br />
• Use of and and or neurons (logic processor)<br />
and<br />
or
<strong>Fuzzy</strong> neural networks:<br />
Example of architectures<br />
• Use of and, or and referential (ref) neurons<br />
ref<br />
and<br />
or
<strong>Fuzzy</strong> cognitive maps<br />
• Representation of concepts and linkages between concepts<br />
• Directed graph: concepts are nodes; linkages are edges<br />
A<br />
+<br />
+<br />
-<br />
-<br />
-<br />
C<br />
D<br />
-<br />
B<br />
• A, B, C, and D = concepts.<br />
• Inhibition (-) or excitation (+) between the concepts (nodes)
Verification and validation<br />
of fuzzy models
Verification and validation<br />
of fuzzy models<br />
• Verification and Validation (V&V) are concerned with the<br />
development of the model and assessment of its usefulness<br />
• Verification is concerned with the analysis of the underlying<br />
processes of constructing the fuzzy model do we follow sound<br />
design principles ?<br />
“Are we building the product right?”<br />
• Validation is concerned with ensuring that the model (product)<br />
meets the requirements of the customer<br />
“Are we building the right product?”
Verification of fuzzy models<br />
• Sound design principles<br />
– iterative development process<br />
– assessment of accuracy<br />
– generalization capabilities<br />
– complexity of the model (Occam’s principle)<br />
– high level of autonomy of the model
<strong>Fuzzy</strong> models: accuracy<br />
• Two ways of expressing accuracy:<br />
– numeric level<br />
– internal level (information granules)
<strong>Fuzzy</strong> models: accuracy<br />
• Numeric level of expressing accuracy<br />
Interface<br />
Interface<br />
x k<br />
y k<br />
target k<br />
Processing<br />
Minimized<br />
error
<strong>Fuzzy</strong> models: accuracy<br />
• Accuracy expressed at the level of fuzzy sets<br />
Interface<br />
Interface<br />
x k<br />
Processing<br />
t k<br />
target k<br />
u k<br />
Minimized<br />
error
Training, validation, and<br />
testing data<br />
• To avoid potential bias in assessment of accuracy, data are split into<br />
– training<br />
– validation<br />
– testing subsets<br />
• Training - testing<br />
– typically 60-40% split<br />
– 10 fold cross-validation (90-10% split)<br />
– leave one out strategy
Validation of fuzzy<br />
models<br />
• Are we building the right model?<br />
• More difficult to quantify:<br />
Interpretability<br />
– transparency of fuzzy models<br />
– stability of the fuzzy model<br />
….<br />
• Very often validation criteria are in conflict<br />
•Curse of dimensionality versus<br />
transparency<br />
Accuracy
Spiral scheme of model development<br />
* incremental design, implementation and testing<br />
* multidimensional space of fundamental characteristics<br />
Stability<br />
Accuracy<br />
Interpretability<br />
Knowledge<br />
Representation
Graph-oriented fuzzy models
<strong>Fuzzy</strong> Models: A General Taxonomy<br />
Source of data/knowledge<br />
single source<br />
Multiple sources<br />
Direction –based<br />
architectures<br />
Graph–oriented<br />
architectures
<strong>Fuzzy</strong> cognitive maps<br />
• Representation of concepts and linkages between concepts<br />
• Directed graph: concepts are represented as nodes;<br />
linkages (associations) are edges<br />
A<br />
+<br />
+<br />
-<br />
-<br />
-<br />
C<br />
D<br />
-<br />
B<br />
• A, B, C, and D = concepts.<br />
• Inhibition (-) or excitation (+) between the concepts (nodes)
Architecture and computing<br />
node-i<br />
w ij<br />
A i<br />
(k+1)= f(<br />
w ik<br />
node-k<br />
node-j<br />
Activation of node<br />
c<br />
∑<br />
A i (k+1)= f( w A (k) + w )<br />
j=<br />
1<br />
ij<br />
j<br />
0i
Architecture and computing:<br />
more computing<br />
Activation of node – nonlinearity with adjustable steepness<br />
Higher-order dynamics: A i (k+1)= F(k, k-1, k-2)…etc
Design: key issues<br />
Nodes – result of abstraction, a collection of<br />
interacting concepts, representation of knowledge<br />
Strength (intensity) of interaction –<br />
+1 , -1 cognitive maps<br />
[-1, 1] fuzzy cognitive maps
Design: key issues<br />
Nodes – result of abstraction, a collection of<br />
interacting concepts, representation of knowledge<br />
INFORMATION GRANULES AND INFORMATION GRANULATION<br />
amplitude =high &<br />
change of amplitude=low<br />
A 1 , A 2 ,…, A c<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
-5 -3 -1<br />
-1<br />
1 3 5<br />
-2<br />
-3<br />
-4<br />
-5<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
-1<br />
0 50 100 150 200 250 300 350 400<br />
-2<br />
-3<br />
-4<br />
-5
Design: key issues<br />
Nodes – result of abstraction, a collection of<br />
interacting concepts, representation of knowledge<br />
Strength (intensity) of interaction –<br />
+1 , -1 cognitive maps<br />
[-1, 1] fuzzy cognitive maps<br />
Learning – parametric optimization
Learning of connections<br />
V =<br />
N c<br />
c<br />
1<br />
∑ ∑||<br />
B (k + 1) − f ( ∑ w<br />
(N −1)c<br />
k=<br />
1<br />
i<br />
i<br />
i= 1<br />
j=<br />
1<br />
ij<br />
A<br />
i<br />
(k) +<br />
w<br />
0i<br />
,σ<br />
i<br />
) ||<br />
2<br />
Parametric optimization<br />
<strong>Fuzzy</strong> clustering<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
-5 -3 -1<br />
-1<br />
1 3 5<br />
-2<br />
-3<br />
-4<br />
-5
Example (1)<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
-1<br />
0 50 100 150 200 250 300 350 400<br />
-2<br />
-3<br />
-4<br />
-5<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
-1<br />
0 50 100 150 200 250 300 350 400<br />
-2<br />
-3<br />
-4<br />
-5<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
-5 -3 -1<br />
-1<br />
1 3 5<br />
-2<br />
-3<br />
-4<br />
-5
Example (2)<br />
(-L, -L)<br />
(-S, -M)<br />
(L, L)<br />
(L, S)<br />
(-S , S)<br />
5<br />
4<br />
∆x k<br />
3<br />
2<br />
1<br />
0<br />
-5 -3 -1<br />
-1<br />
1 3<br />
x k<br />
5<br />
-2<br />
-3<br />
-4<br />
-5
<strong>Fuzzy</strong> cognitive maps:<br />
extensions<br />
A<br />
C<br />
and<br />
or<br />
D<br />
B<br />
E
C<br />
B<br />
D3<br />
<strong>Fuzzy</strong> cognitive maps:<br />
hierarchy of concepts<br />
A<br />
D<br />
or<br />
Level of information granularity<br />
D1<br />
D2
Linguistic models:<br />
from data to granular architectures
Linguistic (granular) modeling:<br />
Main objectives<br />
Direct and constructive usage of information granules<br />
Active role of designer in the formation of focus of the linguistic<br />
model<br />
High interpretability<br />
Reduced learning curve – rapid prototyping with possibilities of<br />
further direct refinements
Clustering and revealed structure vis-àvis<br />
modeling agenda<br />
Clustering (fuzzy clustering) is a direction-free mechanism of<br />
development of information granules, viz. there is no distinction<br />
between independent (input) and dependent (output) objects to<br />
be clustered<br />
One could cluster (group)<br />
(a) Input data x 1 , x 2 , … x N<br />
(b) Concatenated data of input and output [x 1 y 1 ], [x 2 y 2 ],…,<br />
[x N y N ]
Clustering and revealed structure vis-àvis<br />
modeling agenda<br />
Existing drawbacks:<br />
(a) Structure in the input space – clusters could be heterogeneous<br />
in terms of output data<br />
(b) Dimensionality of input space is higher than the output space<br />
formation of proper distance function
Context-based clustering<br />
To align the agenda of fuzzy clustering with the principles of fuzzy<br />
modeling, the following features are considered:<br />
Active role of the designer [customization of the model]<br />
The structural backbone of the model is fully reflective of relationships<br />
between information granules in the input and output space<br />
Clustering : construct clusters in input space X<br />
Context-based Clustering : construct clusters in input space X<br />
given some context expressed in<br />
output space Y
Selected references<br />
W. Pedrycz, Conditional fuzzy C-Means, Pattern Recognition Letters, 17, 1996, 625-<br />
632.<br />
W. Pedrycz, Conditional fuzzy clustering in the design of radial basis function neural<br />
networks, IEEE Trans. on Neural Networks, 9, 1998, 601-612.<br />
W. Pedrycz, A Vasilakos, Linguistic models and linguistic modeling, IEEE Trans. on<br />
Systems, Man and Cybernetics, 29, 1999, 745-757.<br />
W. Pedrycz and K. Kwak, Granular models as a framework of user-centric system<br />
modeling, IEEE Trans. on Systems, Mans, and Cybernetics – Part A, 2006, 727-745.<br />
W. Pedrycz, K.C. Kwak, The development of incremental models, IEEE Trans. on<br />
<strong>Fuzzy</strong> Systems, 15, 3, 2007, 507-518<br />
W. Pedrycz, J. Valente de Oliveira, Development of fuzzy encoding and decoding<br />
through fuzzy clustering, IEEE Transactions on Instrumentation and Measurement,<br />
2008<br />
L. A. Zadeh, <strong>Fuzzy</strong> sets and information granularity, In: M. Gupta, R. Ragade, R. Yager<br />
(Eds.) Advances in <strong>Fuzzy</strong> Set Theory and Applications, Noth-Holland, Amsterdam,<br />
1979, 3-18.
Context-based clustering:<br />
Computing considerations<br />
structure<br />
structure<br />
context<br />
Data<br />
Data
Context-based clustering<br />
Context-based Clustering : construct clusters in input space X<br />
given some context expressed in<br />
output space Y<br />
Context – hint (piece of domain knowledge)<br />
provided by designer who actively impacts the<br />
development of the model
Context-based clustering:<br />
Context design<br />
Context – hint (piece of domain knowledge)<br />
provided by designer who actively impacts the<br />
development of the model. As such, context<br />
is imposed by the designer at the beginning<br />
Realization of context<br />
Designer focus information granule (fuzzy set)<br />
(a) Designer, and (b) clustering of scalar data in output space<br />
Context – fuzzy set (set) formed in the output space
Context-based clustering:<br />
Modeling<br />
Determine structure in input space<br />
given the output is high<br />
Determine structure in input space<br />
given the output is medium<br />
Determine structure in input space<br />
given the output is low<br />
Input space (data)
Context-based clustering:<br />
examples<br />
Find a structure of customer data [clustering]<br />
Find a structure of customer data considering<br />
customers making weekly purchases in the<br />
range [$1,000 $3,000]<br />
no context<br />
context<br />
Find a structure of customer data considering<br />
customers making weekly purchases at the level of<br />
around $ 2,500<br />
Find a structure of customer data considering<br />
customers making significant weekly purchases who<br />
are young<br />
context<br />
context<br />
(compound)
Context-oriented FCM<br />
Data (x k<br />
, target k<br />
), k=1,2,…,N<br />
Contexts: fuzzy sets W 1<br />
, W 2<br />
, …, W p<br />
w jk<br />
= W i<br />
(target k<br />
) membership of j-th context for k-th data<br />
Context-driven partition matrix<br />
U (W )<br />
j<br />
c<br />
N<br />
⎧<br />
= ⎨u<br />
ik<br />
∈[ 0,1]<br />
| ∑ u<br />
ik<br />
= w<br />
jk<br />
∀k<br />
and 0 < ∑ u<br />
⎩<br />
i= 1<br />
k=<br />
1<br />
ik<br />
<<br />
N<br />
⎫<br />
∀i⎬<br />
⎭
Context-oriented FCM:<br />
Optimization flow<br />
Objective function<br />
Q<br />
c<br />
N<br />
= ∑∑<br />
i= 1 k=<br />
1<br />
u<br />
m<br />
2<br />
ik<br />
|| x<br />
k<br />
− v<br />
i<br />
||<br />
Subject to constraint U in U(W j<br />
)<br />
Iterative adjustment of partition matrix and prototypes<br />
u<br />
ik<br />
=<br />
c<br />
∑<br />
j=<br />
1<br />
⎛<br />
⎜<br />
⎜<br />
⎝<br />
x<br />
x<br />
k<br />
k<br />
w<br />
jk<br />
− v<br />
− v<br />
i<br />
j<br />
⎞<br />
⎟<br />
⎟<br />
⎠<br />
2<br />
m−1<br />
v<br />
i<br />
N<br />
∑<br />
k=<br />
1<br />
=<br />
N<br />
∑<br />
k=<br />
1<br />
u x<br />
m<br />
ik<br />
u<br />
m<br />
ik<br />
k
From context-based FCM to<br />
a web of information granules<br />
context<br />
fuzzy set<br />
Input space
From context-based FCM to<br />
a web of information granules<br />
T<br />
1<br />
T<br />
2<br />
T<br />
3<br />
context<br />
fuzzy sets<br />
Input space
Contexts and their design on<br />
a basis of experimental evidence<br />
Selection of triangular fuzzy sets – minimization of reconstruction<br />
error<br />
W i W i+1<br />
m i m i+1<br />
y 0<br />
y
From neurons to<br />
granular neurons<br />
Neurons in neural networks are nonlinear mappings from R n to R<br />
(or [0,1])<br />
Main properties<br />
•Connections are numeric<br />
•Inputs are numeric<br />
•Output is numeric<br />
Granular neurons:<br />
•Connections are information granules (intervals, fuzzy sets…)<br />
•Inputs are information granules (intervals, fuzzy sets…)
Granular neurons<br />
u 1<br />
∑<br />
Y= N(u 1<br />
, u 2<br />
, ..,u c<br />
, A 1<br />
, A 2<br />
, .., A c<br />
) = (Ai ⊗ ui)<br />
A 1<br />
c<br />
i=<br />
1⊕<br />
u 2<br />
Σ<br />
Y<br />
Algebraic operations on<br />
information granules<br />
u c<br />
A c
Granular neurons:<br />
computing (1)<br />
Interval connections A i<br />
= [a i-<br />
, a i+<br />
]<br />
u i<br />
- positive inputs<br />
A ⊗ i<br />
u i<br />
= [a i-<br />
u i<br />
, a i+<br />
u i<br />
]
Granular neurons:<br />
computing (2)<br />
Z i<br />
- fuzzy number<br />
Y(y) = sup {min(Z 1<br />
(y 1<br />
), Z 2<br />
(y 2<br />
), …, Z n<br />
(y n<br />
))}<br />
subject to<br />
y = y 1<br />
+ y 2<br />
+…+y n
Granular neurons:<br />
characteristics<br />
u1=a u 2<br />
= 1-a A 1<br />
= [0.3 3] A 2<br />
= [1.4 7]<br />
α<br />
y
Granular neurons:<br />
characteristics<br />
Connections represented as triangular fuzzy numbers<br />
A i<br />
=<br />
Y<br />
c<br />
=< ∑a<br />
∑ ∑<br />
iui,<br />
miui,<br />
i=<br />
1<br />
c<br />
i=<br />
1<br />
c<br />
i=<br />
1<br />
b<br />
i<br />
u<br />
i<br />
><br />
α<br />
x<br />
α<br />
x
Architecture of granular model<br />
Given a collection of context fuzzy sets W 1<br />
, W 2<br />
, …, W p<br />
(those are provided by designer; active model formation)<br />
Two-phase process:<br />
For each context, construct clusters in input space<br />
(context-based clustering)<br />
Arrange information granules into a web of associations<br />
of information granules formed in the input space and output<br />
space<br />
Rapid prototyping (modeling)
Architecture of granular model<br />
Σ<br />
Σ<br />
Σ<br />
Y<br />
x<br />
Σ<br />
Context-based<br />
clusters<br />
Contexts
Model evaluation:<br />
Performance analysis<br />
Y<br />
Y=<br />
y -<br />
y target y +<br />
Numeric quantification of performance<br />
Granular quantification of performance
Model evaluation:<br />
Numeric quantification<br />
Y<br />
Y=<br />
y target<br />
y -<br />
y +<br />
N<br />
k=<br />
1<br />
2<br />
k k<br />
)<br />
∑(y − target
Model evaluation:<br />
Granular quantification<br />
Y<br />
Y=<br />
y target<br />
y -<br />
y +<br />
N<br />
∑<br />
k=<br />
1<br />
(1−<br />
Y(targetk<br />
))
Enhancements of granular<br />
models<br />
Introduction of bias term to granular neuron<br />
Optimization of contexts – focus of the granular model
Enhancements of granular<br />
models<br />
Introduction of bias term to granular neuron<br />
Σ<br />
Σ<br />
Σ<br />
Y<br />
target k<br />
W 1<br />
W 0<br />
x<br />
W 2<br />
+ y k<br />
y −<br />
y<br />
+<br />
Context-based<br />
clusters<br />
Σ<br />
Contexts<br />
M<br />
M<br />
W p<br />
Σ Y<br />
=< y<br />
−<br />
, y<br />
,<br />
y<br />
+<br />
>
Enhancements of granular<br />
models - bias<br />
M<br />
W 1<br />
W 2<br />
M<br />
W p<br />
W 0<br />
target k<br />
+ y k<br />
y −<br />
y<br />
+<br />
Σ Y<br />
=< y<br />
−<br />
, y<br />
,<br />
y<br />
+<br />
><br />
w<br />
1<br />
N<br />
0<br />
= − ∑ (target<br />
k<br />
− y<br />
k<br />
)<br />
N k=<br />
1<br />
lower bound<br />
p<br />
∑<br />
t=<br />
1<br />
z<br />
t<br />
w<br />
+ t− w<br />
0<br />
modal value ∑ z<br />
tw<br />
p<br />
t=<br />
1<br />
t<br />
+ w<br />
0<br />
upper bound<br />
p<br />
∑<br />
t=<br />
1<br />
z w<br />
t<br />
+ t+ w<br />
0
Enhancements of granular<br />
models – context optimization<br />
T<br />
Conditional<br />
clustering<br />
Context<br />
optimization
Enhancements of granular models –<br />
context optimization (2)<br />
T<br />
max<br />
P<br />
1<br />
N<br />
N<br />
∑ Y( x<br />
k=<br />
1<br />
k<br />
)(target<br />
k<br />
)<br />
P- parameters of contexts<br />
Conditional<br />
clustering<br />
Context<br />
optimization<br />
min P<br />
1<br />
N<br />
N<br />
∑<br />
k=<br />
1<br />
(b<br />
k<br />
− a<br />
k<br />
)
Incremental granular models<br />
Adopting a construct of a linear regression as a first-principle global model, refine<br />
it through granular models that capture remaining and more localized<br />
nonlinearities of the system<br />
fuzzy model = linear regression & local granular models
Incremental granular models<br />
(a)<br />
(b)<br />
(c)
Incremental granular models:<br />
the concept<br />
DATA<br />
{x k , target k }<br />
Linear regression<br />
{x k , e k }<br />
Residuals<br />
Incremental<br />
Granular<br />
model<br />
Granular model
Contexts in incremental granular<br />
models<br />
Context space E<br />
Input space R n
Architecture of the incremental<br />
granular model<br />
x<br />
Linear<br />
Regression<br />
z<br />
Σ<br />
Y<br />
INCREMENTAL<br />
MODEL<br />
Σ<br />
Σ<br />
Σ<br />
E<br />
bias<br />
Σ<br />
Context-based<br />
clustering<br />
fuzzy numbers<br />
(granular information<br />
processed)
Topology of the model<br />
u 11<br />
u 1i<br />
u 1c<br />
Σ<br />
ξ 1<br />
M<br />
u t1<br />
M<br />
W 1<br />
x<br />
u ti<br />
u tc<br />
Σ<br />
ξ t<br />
W t<br />
Σ<br />
E<br />
M<br />
u p1<br />
M<br />
W p<br />
u pi<br />
u pc<br />
Σ<br />
ξ p<br />
Context-based<br />
centers<br />
Contexts<br />
E<br />
=<br />
W<br />
ξ<br />
ξ<br />
1<br />
⊗<br />
1<br />
⊕ W2<br />
⊗<br />
2<br />
⊕....Wn<br />
⊗ξ<br />
n
General design strategy<br />
A. Design of a linear regression in the input – output space, z = L(x; b) with b<br />
denoting a vector of the regression hyperplane, b =[a a 0 ] T . On the basis of<br />
the original data set formed is a collection of input-error pairs, (x k , e k ) where<br />
e k = target-L(x k ,a)<br />
B. Construction of the collection of contexts- fuzzy sets in the space of error of<br />
the regression model E 1 , E 2 , …,E p . Typically triangular fuzzy sets are<br />
considered<br />
C. Context-based FCM completed in the input space and induced by the<br />
individual fuzzy sets of context.<br />
D. Summation of the activation levels of the clusters induced by the<br />
corresponding contexts and their overall aggregation<br />
E. The granular result of the incremental model is affected by eventual bias
Example (1)<br />
“spiky” function<br />
spiky(x)<br />
=<br />
⎧max(x,G(x))<br />
⎨<br />
⎩min(x,<br />
−G(x)<br />
+<br />
2)<br />
if 0 ≤ x ≤ 1<br />
if 1 < x ≤ 2<br />
G(x)<br />
⎛ − (x − c)<br />
exp<br />
⎜<br />
⎝ 2σ<br />
=<br />
2<br />
2<br />
⎟ ⎞<br />
⎠
Example (2)<br />
RMSE =<br />
1<br />
N<br />
N<br />
∑<br />
k=<br />
1<br />
(yk − targetk<br />
)<br />
2<br />
p=c=6<br />
p=c=5<br />
p=c=4<br />
p=c=3
Example (3)<br />
c=p=5, m=2.2
Two-dimensional function
Two-dimensional function:<br />
results<br />
Linear model<br />
Error and distribution of<br />
prototypes
Two-dimensional function:<br />
Contexts and results<br />
1<br />
membership grades<br />
0.8<br />
0.6<br />
0.4<br />
LN MN NZ MP LP<br />
0.2<br />
0<br />
-80 -60 -40 -20 0 20 40<br />
e
Experimental data<br />
Machine Learning and MIS<br />
Number of<br />
contexts and<br />
clusters<br />
RMSE<br />
(Training data)<br />
RMSE<br />
(Test data)<br />
Automobile<br />
MPG<br />
data<br />
Boston<br />
Housing<br />
data<br />
CPU<br />
Performance<br />
data<br />
MIS data<br />
Linear regression 0.194 0.295<br />
Incremental model p=c=6 0.142 0.285<br />
Linear regression 0.240 0.396<br />
Incremental model p=c=6 0.177 0.439<br />
Linear regression 6.544 10.57<br />
Incremental model p=c=6 4.965 11.57<br />
Linear regression 0.626 1.024<br />
Incremental model p=c=6 0.896 0.773
From multitude of perceptions to global<br />
architectures of fuzzy models<br />
Multimodality:<br />
different perspectives:<br />
granularity of information,<br />
variables (features) involved<br />
Different data
<strong>Fuzzy</strong> models:<br />
hierarchical and distributed perspective<br />
granularity/hierarchy<br />
set of features
<strong>Fuzzy</strong> Models: A General Taxonomy<br />
Source of data/knowledge<br />
single source<br />
Multiple sources<br />
Direction –based<br />
architectures<br />
Graph–oriented<br />
architectures
Modeling with for spatio-temporal<br />
data<br />
Model<br />
MODEL-1 MODEL-2 MODEL-3<br />
time<br />
Data-1 Data-2 Data-3<br />
MODEL-1 MODEL-2 MODEL-3<br />
MODEL-1 MODEL-2 MODEL-3<br />
time<br />
Data-1 Data-2 Data-3<br />
Data-1 Data-2 Data-3<br />
time
Two general development strategies<br />
SELECTION OF A “MEANINGFUL” SUBSET OF<br />
INFORMATION GRANULES
Two general development strategies<br />
(1) HIERARCHICAL DEVELOPMENT OF INFORMATION<br />
GRANULES (INFORMMATION GRANULES OF HIGHER<br />
TYPE)<br />
Information granules<br />
Type -2<br />
Information granules<br />
Type -1
Two general development strategies<br />
(2) HIERARCHICAL DEVELOPMENT OF INFORMATION<br />
GRANULES AND THE USE OF VIEWPOINTS<br />
Information granules<br />
Type -2<br />
viewpoints<br />
Information granules<br />
Type -1
Two general development strategies<br />
(3) HIERARCHICAL DEVELOPMENT OF INFORMATION<br />
GRANULES – A MODE OF SUCCESSIVE CONSTRUCTION
Main design phases<br />
Φi[1]<br />
Ai[1]<br />
vi[1]<br />
F1<br />
Determination of information granules<br />
Fii<br />
data<br />
Φi[ii]<br />
Ai[ii]<br />
vi[ii]<br />
Φi[p]<br />
Ai[p]<br />
zi<br />
z1<br />
zc<br />
Construction of associated local<br />
models<br />
vi[p]<br />
Fp<br />
F
A global granular view:<br />
choosing representative information granules<br />
Φ i[1]<br />
A i[1]<br />
v i[1]<br />
F 1<br />
prototypes z 1<br />
, z 2<br />
, …, z c<br />
data<br />
Φ i[ii]<br />
A i[ii]<br />
F ii<br />
heterogeneous space<br />
v i[ii]<br />
Recall “ standard” expression:<br />
z i<br />
z c<br />
Φ i[p]<br />
A i[p]<br />
v i[p]<br />
F p<br />
z 1<br />
F<br />
A<br />
i<br />
( x)[ii]<br />
=<br />
c[ii]<br />
∑<br />
j=<br />
1<br />
⎛<br />
⎜<br />
|| x − v<br />
⎝ || x − v<br />
1<br />
i<br />
[ii] || ⎞<br />
⎟<br />
[ii] ||<br />
⎠<br />
j<br />
2/(m −1)<br />
ii
Information granules and<br />
their representatives<br />
Represent v k<br />
[ii] with the use of z 1<br />
, z 2<br />
, …, z c<br />
u i<br />
(v k<br />
[ii]) =<br />
c<br />
∑<br />
j=1<br />
1<br />
⎛ || v k<br />
[ii]− z i<br />
|| Fii ∩F<br />
⎜<br />
⎝ || v k<br />
[ii]− z j<br />
|| Fii ∩F<br />
⎞<br />
⎟<br />
⎠<br />
2/(m−1)<br />
z 1<br />
z 2<br />
v 1 [ii]<br />
z c<br />
F<br />
F ii
Representation of fuzzy sets:<br />
two performance measures<br />
Entropy measure<br />
Reconstruction criterion (error)
Expressing performance through<br />
entropy measure<br />
p<br />
∑<br />
ii=<br />
1<br />
c<br />
∑<br />
i=<br />
1<br />
c[ii]<br />
∑<br />
k=<br />
1<br />
H(u<br />
i<br />
( v<br />
k<br />
[ii]))
Reconstruction error<br />
Q =<br />
p<br />
∑<br />
ii=<br />
1<br />
c[ii]<br />
∑<br />
k=<br />
1<br />
2<br />
|| v ˆ( v<br />
k[ii])<br />
− v<br />
k[ii] ||<br />
F<br />
ii<br />
where<br />
vˆ<br />
( v<br />
k<br />
[ii])<br />
c<br />
∑<br />
m<br />
m<br />
= u ( v [ii]) z vˆ<br />
( v [ii]) = u ( v [ii]) z / u ( v [ii])<br />
i=<br />
1<br />
m<br />
i<br />
k<br />
i<br />
k<br />
c<br />
∑<br />
i=<br />
1<br />
i<br />
k<br />
i<br />
c<br />
∑<br />
i=<br />
1<br />
i<br />
k<br />
Requirement of “coverage” condition<br />
c<br />
U<br />
k=<br />
1<br />
F<br />
p<br />
U<br />
= F<br />
i k<br />
i=<br />
1<br />
i
Optimization problem<br />
Form a collection of prototypes Z = {z 1<br />
, z 2<br />
, …, z c<br />
} such that<br />
entropy (or reconstruction error)<br />
is minimized while satisfying coverage criterion<br />
c<br />
U<br />
k=<br />
1<br />
F<br />
p<br />
U<br />
= F<br />
i k<br />
i=<br />
1<br />
i<br />
Min Z<br />
Q subject to<br />
c<br />
U<br />
k=<br />
1<br />
F<br />
p<br />
U<br />
= F<br />
i k<br />
i=<br />
1<br />
i<br />
Optimization of fuzzification coefficient (m)<br />
Min Z<br />
Q subject to m>1 and<br />
c<br />
U<br />
k=<br />
1<br />
F<br />
p<br />
U<br />
= F<br />
i k<br />
i=<br />
1<br />
i
Design of local models<br />
Φi[1]<br />
Ai[1]<br />
vi[1]<br />
F1<br />
data<br />
Φi[ii]<br />
Ai[ii]<br />
vi[ii]<br />
Fii<br />
zi<br />
zc<br />
Φi[p]<br />
Ai[p]<br />
vi[p]<br />
Fp<br />
z1<br />
F
Knowledge sharing<br />
knowledge<br />
data<br />
Structure<br />
Model (e.g., fuzzy rule-based)<br />
Predictor<br />
Decision-making strategy<br />
…..<br />
Signature of knowledge<br />
phenomenon, process, system…<br />
Structure – clusters [prototypes]<br />
Model: condition and conclusion parts<br />
- if x is A i<br />
then y is f i<br />
(x, a i<br />
)
Knowledge sharing and<br />
collaboration<br />
knowledge<br />
data-2<br />
Structure<br />
Model (e.g., fuzzy rule-based)<br />
Predictor<br />
Decision-making strategy<br />
…..<br />
Signature of knowledge<br />
knowledge<br />
data-1<br />
Structure<br />
Model (e.g., fuzzy rule-based)<br />
Predictor<br />
Decision-making strategy<br />
…..<br />
Signature of knowledge<br />
knowledge<br />
Structure<br />
Model (e.g., fuzzy rule-based)<br />
Predictor<br />
Decision-making strategy<br />
…..<br />
data-P<br />
Signature of knowledge<br />
phenomenon, process, system…
Knowledge sharing and<br />
collaboration<br />
knowledge<br />
data-2<br />
Structure<br />
Model (e.g., fuzzy rule-based)<br />
Predictor<br />
Decision-making strategy<br />
…..<br />
Signature of knowledge<br />
knowledge<br />
data-1<br />
Structure<br />
Model (e.g., fuzzy rule-based)<br />
Predictor<br />
Decision-making strategy<br />
…..<br />
Signature of knowledge<br />
knowledge<br />
Structure<br />
Model (e.g., fuzzy rule-based)<br />
Predictor<br />
Decision-making strategy<br />
…..<br />
data-P<br />
Signature of knowledge
Collaborative<br />
structure development (1)<br />
Information<br />
granules<br />
data-1<br />
data-2<br />
data-P<br />
phenomenon, process, system…
Collaborative<br />
structure development (2)<br />
Information<br />
granules of<br />
higher type<br />
Information<br />
granules<br />
data-1<br />
data-2<br />
data-P<br />
phenomenon, process, system…
Collaborative clustering<br />
Information<br />
Information<br />
granules Information<br />
granules<br />
granules<br />
data-1 data-1 data-1<br />
data-2 data-2 data-2<br />
phenomenon, phenomenon, process, process, system…<br />
phenomenon, process, system… system…<br />
data-P data-P data-P<br />
Discover a structure in a collaborative fashion by communicating<br />
findings produced at the level of local data sites.<br />
Exchange of findings in the form of information granules<br />
constructed for each data site<br />
Further usage of such findings in refining and directing (navigating)<br />
search in local data
Collaborative structure determination:<br />
Information granules of higher order<br />
Information<br />
Information<br />
granules of<br />
granules higher type of<br />
higher type<br />
Prototypes<br />
(higher order)<br />
Clustering<br />
Information<br />
Information<br />
granules<br />
granules<br />
data-1<br />
data-1<br />
data-2<br />
data-2<br />
phenomenon, process, system…<br />
phenomenon, process, system…<br />
data-P<br />
data-P<br />
prototypes<br />
D[1] D[2] D[P]
Determining correspondence between<br />
clusters<br />
Prototypes<br />
(higher order)<br />
z j<br />
Clustering<br />
Select prototypes in D[1], D[2], …, D[p] associated with z j<br />
with the highest degree of membership
Determining correspondence between<br />
clusters<br />
z j<br />
D[ii]<br />
v i<br />
[ii]<br />
λ<br />
ij<br />
[ii]<br />
=<br />
c[ii]<br />
∑<br />
k=<br />
1<br />
⎛ || v<br />
⎜<br />
⎝<br />
|| v<br />
i<br />
k<br />
1<br />
[ii] − z<br />
[ii] − z<br />
j<br />
j<br />
|| ⎞<br />
⎟<br />
||<br />
⎠<br />
2<br />
Prototype i 0<br />
associated with prototype z j<br />
λ<br />
[ii]<br />
=<br />
max<br />
i0 j<br />
i=<br />
1,2,...,c[ii]<br />
λ<br />
ij
Family of associated prototypes<br />
Prototype i 1<br />
in D[1] associated with prototype z j<br />
Prototype i 2<br />
Prototype i p<br />
in D[2] associated with prototype z j<br />
…<br />
in D[p] associated with prototype z j<br />
v<br />
i<br />
1<br />
[1],<br />
v<br />
i<br />
2<br />
[2],....,<br />
v<br />
i<br />
p<br />
[P]<br />
λ<br />
i<br />
1<br />
,<br />
λ<br />
i<br />
2<br />
,....,<br />
λ<br />
i<br />
p
From numeric prototypes to<br />
granular prototypes<br />
v<br />
i<br />
1<br />
[1],<br />
v<br />
i<br />
2<br />
[2],....,<br />
v<br />
i<br />
p<br />
[P]<br />
λ<br />
i<br />
1<br />
,<br />
λ<br />
i<br />
2<br />
,....,<br />
λ<br />
i<br />
p<br />
individual coordinate of the associated prototypes:<br />
a 1<br />
a 2<br />
…. a p<br />
µ 1<br />
µ 2<br />
…. µ p<br />
R<br />
[0,1]<br />
Information granule
The principle of justifiable granularity:<br />
Interval representation<br />
a 1<br />
a 2<br />
…. a p<br />
1<br />
µ 1<br />
µ 2<br />
…. µ p<br />
0<br />
b<br />
a 0<br />
d<br />
if a i<br />
∈ [b,d] then elevate to membership grades to 1<br />
required change : 1- µ i
The principle of justifiable granularity:<br />
Interval representation<br />
a 1<br />
a 2<br />
…. a p<br />
1<br />
µ 1<br />
µ 2<br />
…. µ p<br />
0<br />
b<br />
a 0<br />
d<br />
if a i<br />
∉ [b,d] then reduce membership grades to 0<br />
required change: µ i
The principle of justifiable granularity:<br />
optimization criterion<br />
1<br />
0<br />
z 1<br />
z 2<br />
∑<br />
a i ∈[b,d]<br />
∑<br />
a i ∉[b,d]<br />
Min b,d ∈R:b≤d<br />
{ (1− µ i<br />
) + µ i<br />
}
Hyperbox prototypes<br />
i ≠<br />
j:<br />
H<br />
i<br />
∩ H<br />
j<br />
=<br />
∅<br />
(the number of<br />
H i<br />
H j<br />
level)<br />
clusters at the aggregation
Interval-valued fuzzy sets<br />
and granular prototypes<br />
x<br />
H i<br />
H j
Interval-valued fuzzy sets<br />
and granular prototypes<br />
v i<br />
x<br />
||<br />
x − v<br />
i<br />
||<br />
min<br />
|<br />
x<br />
−<br />
v<br />
i<br />
||<br />
max<br />
Bounds of distances computed coordinate-wise
Interval-valued fuzzy sets:<br />
membership function<br />
∑<br />
∑<br />
=<br />
−<br />
−<br />
=<br />
−<br />
+<br />
⎟<br />
⎟<br />
⎠<br />
⎞<br />
⎜<br />
⎜<br />
⎝<br />
⎛<br />
−<br />
−<br />
=<br />
⎟<br />
⎟<br />
⎠<br />
⎞<br />
⎜<br />
⎜<br />
⎝<br />
⎛<br />
−<br />
−<br />
=<br />
c<br />
1<br />
j<br />
1<br />
2<br />
min<br />
j<br />
max<br />
i<br />
i<br />
c<br />
1<br />
j<br />
1<br />
2<br />
max<br />
j<br />
min<br />
i<br />
i<br />
||<br />
||<br />
||<br />
||<br />
1<br />
)<br />
(<br />
u<br />
||<br />
||<br />
||<br />
||<br />
1<br />
)<br />
(<br />
u<br />
m<br />
m<br />
v<br />
x<br />
v<br />
x<br />
x<br />
v<br />
x<br />
v<br />
x<br />
x<br />
Upper bound<br />
Lower bound
Collaborative structure determination:<br />
Structure refinement<br />
Information<br />
Information<br />
granules of<br />
granules of<br />
higher type<br />
higher type<br />
Information<br />
Information<br />
granules<br />
granules<br />
data-1<br />
data-1<br />
data-2<br />
data-2<br />
data-P<br />
data-P<br />
Feedback<br />
and structure<br />
refinement<br />
phenomenon, process, system…<br />
phenomenon, process, system…
Collaborative structure determination:<br />
Structure refinement<br />
Information<br />
Information<br />
granules of<br />
granules higher of type<br />
higher type<br />
Information<br />
Information<br />
granules<br />
granules<br />
data-1<br />
data-1<br />
data-2<br />
data-2<br />
phenomenon, process, system…<br />
phenomenon, process, system…<br />
data-P<br />
data-P<br />
Iterate<br />
Clustering at the local level<br />
Sharing findings and clustering at the higher (global) level<br />
Assessment of quality of clusters in light of the global structure<br />
γ i (U)[ii]<br />
Refinement of clustering<br />
Q[ii] =<br />
c[ii]<br />
2<br />
∑ ∑ γi (U)[ii]|| xk<br />
− vi[ii]<br />
||<br />
i = 1 x ∈X[ii]<br />
k<br />
Until termination criterion satisfied
Towards enhanced interpretability of<br />
fuzzy models<br />
<strong>Fuzzy</strong> model
Towards enhanced interpretability of<br />
results of decision support models<br />
User<br />
Qualitative membership degrees<br />
A(x)=high<br />
R(A, B)=medium<br />
Numeric membership degrees<br />
A(x)=0.87<br />
<strong>Fuzzy</strong><br />
model<br />
R(A, B)=0.51
Interpretability of fuzzy sets through<br />
type-2 fuzzy sets<br />
H M L<br />
Linguistic descriptors of membership<br />
Numeric membership degrees<br />
linguistic<br />
quantification<br />
x
Interpretability of fuzzy sets through<br />
type-2 fuzzy sets<br />
H M L<br />
A<br />
L L L M H H M M M L L<br />
linguistic<br />
quantification<br />
x<br />
x<br />
B 1<br />
, B 2<br />
, …, B r<br />
defined in [0,1] – treated as qualitative evaluators<br />
The least uncertain representation of A in terms of {B 1<br />
, B 2<br />
, …, B r<br />
}<br />
V<br />
=<br />
∫∑<br />
X<br />
r<br />
i=<br />
1<br />
H(B i (A(x)))dx<br />
Min
Towards enhanced interpretability of<br />
fuzzy models<br />
User<br />
Interpretability layer<br />
<strong>Fuzzy</strong> model<br />
Information<br />
granules<br />
Data/<br />
environment