195.130.93.18195.130.93.18/vgkabs/documents/sj252012kaburlasosp.pdf1 a lattice-computing ensemble...
TRANSCRIPT
1
A Lattice-Computing Ensemble for Reasoning
Based on Formal Fusion of Disparate Data Types,
and an Industrial Dispensing Application
Vassilis G. Kaburlasos and Theodore Pachidis
Department of Industrial Informatics
Technological Educational Institution of Kavala
65404 Kavala, Greece
Emails: {vgkabs,pated}@teikav.edu.gr
Abstract1
By “fusion” this work means integration of disparate types of data including (intervals of) real numbers as2
well as possibility/probability distributions defined over the totally-ordered lattice(R,≤) of real numbers. Such3
data may stem from different sources including (multiple/multimodal) electronic sensors and/or human judgement.4
The aforementioned types of data are presented here as different interpretations of a single data representation,5
namely Intervals’ Number (IN). It is shown that the setF of INs is a partially-ordered lattice(F,¹) originating,6
hierarchically, from(R,≤). Two sound, parametricinclusion measurefunctionsσ : FN × FN → [0, 1] result in the7
Cartesian product lattice(FN ,¹) towards decision-making based on reasoning. In conclusion, the space(FN ,¹)8
emerges as a formal framework for the development of hybrid intelligent fusion systems/schemes. A fuzzy lattice9
reasoning (FLR) ensemble scheme, namelyFLR pairwise ensemble, or FLRpefor short, is introduced here for sound10
decision-making based on descriptive knowledge (rules). Advantages include the sensible employment of a sparse11
rule base, employment of granular input data (to cope with imprecision/uncertainty/vagueness), and employment of12
all-order data statistics. The advantages as well as the performance of our proposed techniques are demonstrated,13
comparatively, by computer simulation experiments regarding an industrial dispensing application.14
Index Terms15
Disparate Data Fusion, Ensemble of Experts, Fuzzy lattice reasoning (FLR), Granular data, Inclusion measure,16
Intervals’ number (IN), Lattice-computing, Lattice theory, Sparse rules17
2
I. I NTRODUCTION18
In the domain ofSoft Computingor, equivalently,Computational Intelligence, the term “hybrid (system/algo-19
rithm)” frequently denotes an integration of different techniques/technologies including artificial neural networks,20
fuzzy systems, evolutionary/swarm computing, etc. towards improving an index of performance in real-world21
applications [1], [15]; the term “intelligence” is pertinent to decision-making, e.g. in pattern classification/recognition22
[81]; moreover, the term “(intelligent) fusion” may signify an aggregate intelligence towards improving decision-23
making [47]. In the aforementioned sense, a “hybrid intelligent fusion system” may be a Multiple Classifier System24
(MCS) [45], [48] also known in the literature asClassifier Ensemble[16], [58], [64], Committee[21], [79], or25
Voting Consensus[5], [50]. Note that a number of MCS architectures/strategies including applications have been26
reported [22], [28], [29], [46], [49], [51], [54], [55], [69], [70], [73], [80], [84], [85]. The MCS techniques are,27
typically, of statistical nature [33] in the Euclidean spaceRN . Nevertheless, a “hybrid intelligent fusion system”28
may be defined otherwise, as explained next.29
The term “fusion” may, alternatively, denote an integration of data stemming from multiple, even heterogeneous,30
sources including (multimodal) electronic devices as well as human judgement [6], [9], [13], [17], [20], [26], [52],31
[56], [63], [65], [67]. In the latter context, there is a keen interest in formal frameworks for unified decision-making32
based on disparate types of data that may accommodate uncertainty [9], [18], [78]. One such a framework has33
been proposed lately [35], in an information engineering context, based on mathematicallattice theoryas follows.34
Different authors have recognized that several types of data of practical interest, includinggranules[61], [83],35
are partially(lattice)-ordered [37], [71]. Hence, lattice theory emerged as a formal framework for the fusion of36
disparate data types [35]. In such context,fuzzy lattice reasoning(FLR) was originally proposed [36], [41], [43]37
as a specific rule-based scheme for classification in a complete lattice(L,¹) data domain including, as a special38
case, the lattice of hyperboxes in the Euclidean spaceRN . In this work, FLR (reasoning) is defined, more widely,39
as any employment of aninclusion measurefunction σ : L × L → [0, 1] for decision-making. Therefore, in the40
context of this work, the term “intelligent” is pertinent to “(FLR) reasoning”.41
Instead of a general mathematical lattice this work considers a specific one originating hierarchically from42
the totally-ordered lattice(R,≤) of real numbers. Note that the latter (lattice) has stemmed, historically, from43
the conventional measurement process of successive comparisons [35], [41]. Our interest in lattice(R,≤) was44
motivated by the existence of vast quantities of real number measurements stored worldwide. Therefore, we sought45
convenient data/information representations based onR. Hence, the complete lattice(F,¹) of Intervals’ Numbers46
(IN) emerged, as detailed below, where a IN is a unified data representation including real numbers, intervals, and47
probability/possibility distributions [59]. In conclusion, the Cartesian product lattice(FN ,¹) is introduced here as48
a formal framework for developing hybrid intelligent fusion systems/schemes, where an element of lattice(FN ,¹)49
is interpreted here as either a rule (of a FLR scheme) or as an input to a FLR scheme.50
In previous work, a FLR scheme for classification has been implemented on theσ-FLNMAP neural network51
architecture [35], [42], [44]. Note that the latter (neural network architecture) was introduced as an enhancement of52
3
the fuzzy-ARTMAP, or FAM for short, neural classifier [11]. More specifically, theσ-FLNMAP has extended the53
applicability domain of FAM from the lattice of hyperboxes inRN to any complete lattice data domain. Moreover,54
even in the Euclidean spaceRN , that is FAM’s sole “applicability domain”, classifierσ-FLNMAP has demonstrated55
significant improvements including tunable nonlinearities as well as the capacity to deal with both nonoverlapping56
hyperboxes and granular (hyperbox) input data [35], [42].57
Due to the fact that both classifiersσ-FLNMAP and FAM areunstable, in the sense that their testing accuracy58
depends on the order of presenting the training data [19], [42], it turns out that both of them make good candidates59
for Voting classification schemes [10], [35], [68]. Indeed, empirical studies have clearly demonstrated an improved60
testing accuracy as well as a more stable testing accuracy for both FAM [3], [12], [60] andσ-FLNMAP [35],61
[44] in RN . Later work has extended the applicability ofσ-FLNMAP from the lattice of hyperboxes to the lattice62
(F,¹) of INs based on FLR [41]. In all, FLR is aLattice-Computingscheme as explained next.63
Lattice-Computing (LC) is a term introduced by Grana [23] to denote any computation in a mathematical lattice.64
Grana and colleagues have demonstrated a number of LC techniques in signal/image processing applications [24],65
[25]. In particular, they have employed mathematical morphology techniques in the totally-ordered lattice of real66
numbers. It turns out that FLR is also a LC scheme, in particular for reasoning, as shown below.67
This paper is based on previously published work on FLR. The novelties of this work include the following.68
First, it presents a space of INs as a formal information fusion framework including a large number of references as69
well as pertinent discussions; a novel mathematical proof is also presented here. Second, it includes mathematical70
notation improvements. Third, it introduces an enhanced definition of FLR. Fourth, it demonstrates the “in principle”71
accommodation of granular inputs. Fifth, it introduces a novel decision-making scheme, that is a descriptive (rule-72
based) FLR ensemble of experts. Sixth, it shows a number of illustrative, new examples including figures. Seventh,73
it demonstrates preliminary (computer simulation) results regarding an industrial application.74
The layout of this work is as follows. Section II presents a formal framework for fusion/integration of disparate75
data types. Section III describes our proposed FLR ensemble scheme. Section IV outlines an industrial application.76
Section V demonstrates, comparatively, preliminary experimental results. Section VI concludes by summarizing77
our contribution. The Appendix presents novel mathematical notation as well as a novel mathematical proof.78
II. A F ORMAL INFORMATION FUSION FRAMEWORK79
This section introduces constructively, in four steps, a formal information fusion framework, namely the80
Cartesian product lattice(FN ,¹) of Intervals’ Numbers (INs). Different interpretations of INs are also presented.81
Note that the four-level hierarchy of lattices presented here is a novelty of this work. For the interested reader,82
useful notions and tools regarding lattice theory are summarized in the Appendix.83
4
A. The Complete Lattice (R,≤)84
The setR of real numbers is a totally-ordered, non-complete lattice denoted by(R,≤). It turns out that(R,≤)85
can be extended to a complete lattice by including both symbols “−∞” and “+∞”. In conclusion, the complete86
lattice (R,≤) emerges, whereR = R∪{−∞, +∞}. Note that previous work has, erroneously, assumed that lattice87
(R,≤) is complete [37], [59]. Even though the aforementioned error is not critical, this work considers, instead,88
the complete lattice(R,≤) 1. We remark that complete lattices are important not only in defining aninclusion89
measurefunction, as shown in the Appendix, but they are also important inmathematical morphology[57], [66].90
On the one hand, any strictly increasing functionv : R → R is a positive valuation in the complete lattice91
(R,≤). Motivated by the two constraints presented in the Appendix (subsection B), here we consider positive92
valuation functionsv : R → R≥0 such that bothv(−∞) = 0 andv(+∞) < +∞. On the other hand, anybijective93
(i.e. one-to-one), strictly decreasing functionθ : R → R is a dual isomorphic function in lattice(R,≤). We will94
refer to functionsθ(.) and v(.) simply asdual isomorphicand positive valuation, respectively. Useful extensions95
to the corresponding lattice of intervals are presented next.96
B. The Complete Lattice (∆,¹) Induced from (R,≤)97
A generalized intervalis defined in lattice(R,≤) as follows.98
Definition 1: Generalized intervalis an element of the product lattice(R,≤∂)× (R,≤).99
Recall that≤∂ in Definition 1 denotes thedual (i.e. converse) of order relation≤ in lattice (R,≤), i.e.≤∂≡≥.100
Product lattice(R,≤∂)× (R,≤) ≡ (R× R,≥ × ≤) will be denoted, simply, by(∆,¹).101
A generalized interval will be denoted by[x, y], wherex, y ∈ R. It follows that themeet(f) and join (g) in102
lattice (∆,¹) are given, respectively, by[a, b] f [c, d] = [a ∨ c, b ∧ d] and [a, b] g [c, d] = [a ∧ c, b ∨ d].103
The set ofpositive (negative) generalized intervals[a, b], characterized bya ≤ b (a > b), is denoted by∆+104
(∆−). It turns out that(∆+,¹) is a poset, namelyposet of positive generalized intervals. Note that poset(∆+,¹)105
is isomorphicto the poset(τ(R),¹) of conventional intervals (sets) inR, i.e. (τ(R),¹) ∼= (∆+,¹). We augmented106
poset(τ(R),¹) by a least (empty) interval, denoted byO = [+∞,−∞] – We remark that agreatestinterval107
I = [−∞,+∞] already exists inτ(R). Hence, the complete lattice (τO(R) = τ(R) ∪ {O},¹)∼= (∆+ ∪ {O},¹)108
emerged. In the sequel, we will employ isomorphic lattices(∆+ ∪ {O},¹) and (τO(R),¹), interchangeably. We109
point out that a trivial interval[x, x] is anatomin the complete lattice(τO(R),¹), where an atom[x, x] by definition110
satisfies both[+∞,−∞] = O ≺ [x, x] and there is no interval[a, b] ∈ (τO(R),¹) such thatO ≺ [a, b] ≺ [x, x].111
Consider both a positive valuation functionv : R → R≥0 and a dual isomorphic functionθ : R → R. Then,112
proposition 6.2 (in the Appendix) implies that functionv∆ : ∆ → R given by v∆([a, b]) = v(θ(a)) + v(b) is a113
1Personal communication with Peter Sussner in the context of the Hybrid Artificial Intelligence Systems (HAIS ’2010) InternationalConference, 23-25 June 2010, San Sebastian, Spain. It is understood that the authors here assume full responsibility for possible errors.
5
positive valuation in lattice(∆,¹). There follow bothv∆(O = [+∞,−∞]) = 0 andv∆(O = [−∞, +∞]) < +∞.114
Therefore, based on Theorem 6.1 (in the Appendix), the following two inclusion measures emerge in lattice(∆,¹).115
(1) σf([a, b] ¹ [c, d]) = v(θ(a∨c))+v(b∧d)v(θ(a))+v(b) , and116
(2) σg([a, b] ¹ [c, d]) = v(θ(c))+v(d)v(θ(a∧c))+v(b∨d) .117
The above inclusion measures are extended to the lattice(τO(R),¹) of intervals (sets) as follows.118
(1) σf([a, b] ¹ [c, d]) = v(θ(a∨c))+v(b∧d)v(θ(a))+v(b) , if a ∨ c ≤ b ∧ d; otherwise,σf([a, b] ¹ [c, d]) = 0, and119
(2) σg([a, b] ¹ [c, d]) = v(θ(c))+v(d)v(θ(a∧c))+v(b∨d) .120
Functionsθ(.) andv(.) can be selected in different ways; for instance, choosingθ(x) = −x andv(.) such that121
v(x) = −v(−x) it follows v∆([a, b]) = v(b)− v(a). Here, we select a pair of parametric functionsv(x) andθ(x)122
so as to satisfy equalityv∆([x, x]) = v(θ(x)) + v(x) = Constant required for atoms by a popular FLR algorithm123
[42], [43]. Eligible pairs of functionsv(x) and θ(x) include, first, v(x) = A1+e−λ(x−µ) and θ(x) = 2µ − x,124
where A, λ ∈ R≥0, µ, x ∈ R and, second,v(x) = px and θ(x) = Q − qx, where p, q,Q > 0, x ∈ [0, A].125
Since it was assumedv(θ(x)) + v(x) = Constant, for the latter pair of functionsv(x) and θ(x) it follows126
v(θ(x)) + v(x) = p[Q + (1− q)x] = Constant; therefore,q = 1.127
C. The Complete Lattice(F,¹) Induced from (∆,¹)128
Based on generalized interval analysis above, this subsection presentsintervals’ numbers(INs). A more129
general number type is defined in the first place, next.130
Definition 2: Generalized interval number, or GIN for short, is a functionG : (0, 1] → ∆.131
Let G denote the set of GINs. It follows complete lattice(G,¹), as the Cartesian product of complete lattices132
(∆,¹). Our interest here focuses on thesublattice2 of intervals’ numbersdefined next.133
Definition 3: An Intervals’ Number, or IN for short, is a GINF such that bothF (h) ∈ (∆+ ∪ {O}) and134
h1 ≤ h2 ⇒ F (h1) º F (h2).135
Let F denote the set of INs. It follows that(F,¹) is a complete lattice with least elementO = O(h) =136
[+∞,−∞], h ∈ (0, 1] and greatest elementI = I(h) = [−∞,+∞], h ∈ (0, 1]. Conventionally, a IN will be137
denoted by a capital letter in italics, e.g.F ∈ F.138
Definition 3 implies that a INF is a function from interval(0, 1] to the setτ(R)∪ {[+∞,−∞]} of intervals,139
i.e. F (h) = [ah, bh], h ∈ (0, 1], where both interval-endsah andbh are functions ofh ∈ (0, 1].140
The following two inclusion measures emerge, respectively, in the complete lattice (F,¹) of INs [34], [35]:141
(1) σf(F1 ¹ F2) =1∫0
σf(F1(h) ¹ F2(h))dh.142
(2) σg(F1 ¹ F2) =1∫0
σg(F1(h) ¹ F2(h))dh.143
2A sublatticeof a lattice(L,¹) is another lattice(S,¹) such thatS ⊆ L.
6
The following Proposition derives from [37].144
Proposition 2.1: Consider a continuousdual isomorphicfunction θ : R → R and a continuouspositive145
valuation function v : R → R≥0. Let X0(h) = [x0, x0], h ∈ (0, 1] be a trivial (point) IN, moreover letE(h),146
h ∈ (0, 1] be a IN withupper-semicontinuousmembership functionmE : R → R. Thenσf(X0 ¹ E) = mE(x0).147
We remark that Proposition 2.1 couples a IN’s two different representations, namely theinterval-representation148
and themembership-function-representation. The principal advantage of the former (interval) representation is149
that it enables useful algebraic operations, whereas the principal advantage of the latter (membership function)150
representation is that it enables convenient interpretions, e.g. fuzzy logic interpretions, etc.151
D. Extensions to More Dimensions152
A N -tuple IN will be denoted by a capital letter in bold, e.g.F = (F1, ..., FN ) ∈ FN . Lattice (FN ,¹) is153
the “fourth level” in a hierarchy of complete lattices whose “first level”, “second level” and “third level” include154
lattices (R,≤), (∆,¹) and (F,¹), respectively.155
The following Proposition derives from [37].156
Proposition 2.2: ConsiderN complete lattices(Li,¹), i ∈ {1, ..., N} each one equipped with an inclusion157
measure functionσi : Li× Li → [0, 1], respectively. ConsiderN -tuplesx = (x1, . . . , xN ) andy = (y1, . . . , yN ) in158
L = L1×· · ·×LN . Furthermore, consider the conventional lattice orderingx ¹ y ⇔ xi ¹ yi, ∀i ∈ {1, ..., N}. Then,159
both functions (1)σ∧ : L× L → [0, 1] given byσ∧(x ¹ y) = mini∈{1,...,N}
{σi(xi ¹ yi)}, and (2)σΠ : L× L → [0, 1]160
given byσΠ(x ¹ y) = Πi∈{1,...,N}
σi(xi ¹ yi), are inclusion measures in lattice(L,¹).161
We remark that Propositions 2.1 and 2.2 establish that, for trivial inputs, an inclusion measure reduces to162
standard fuzzy inference system (FIS) practices [37].163
E. IN Interpretations, Representation Issues & More, Useful Results164
The complete lattice(F,¹) of INs has been studied in a series of publications [34], [38], [40], [41], [59],165
[62]. In particular, it has been shown that a IN is a mathematical object, which may admit different interpretations166
as follows. First, based on the “resolution identity theorem” [82], a INF (h), h ∈ (0, 1] may be interpreted as167
a fuzzy number, whereF (h) is the correspondingα-cut for α = h. Hence, a INF : (0, 1] → τO(R) may,168
equivalently, be represented by anupper-semicontinuousmembership functionmF : R → (0, 1] – Note that a169
number of authors have employedα-cuts and/or intervals in fuzzy logic applications [2], [74], [75], [76], [77].170
There follows equivalencemF1(x) ≤ mF2(x) ⇔ F1(h) ¹ F2(h), wherex ∈ R, h ∈ (0, 1] [59]. Second, a IN171
F (h), h ∈ (0, 1] may also be interpreted as a probability distribution such that intervalF (h) includes100(1−h)%172
of the distribution, whereas the remaining100h% is split even both below and above intervalF (h).173
Fig.1 explains how a IN can be constructed from a population of (real number) data samples using algorithm174
CALCIN [34], [35], [39], [59], [62]. More specifically, Fig.1(a) displays the data itself. Fig.1(b) displays a histogram175
of the data in Fig.1(a) in10 steps of length∆x = 0.04. Hence, the histogram of Fig.1(b) may be thought of as176
7
a probability density function(pdf) approximation, which (histogram) asymptotically tends to the corresponding177
pdf when both∆x → 0 and the number of data samples tends to infinity. Fig.1(c) displays the corresponding178
cumulative distribution function (PDF). Finally, Fig.1(d) displays a IN computed from the PDF of Fig.1(c) using179
the algebraic formulas shown within Fig.1(d); that is, algorithm CALCIN.180
Fig.2 shows the two different representations of the IN (F ) computed in Fig.1(d). More specifically, Fig.2(a)181
displays the membership-function-representation of INF , whereas Fig.2(b) displays the corresponding interval-182
representation forL = 32 different levels spaced evenly over the interval(0, 1]. Triangular INs are of particular183
significance in practice, therefore they are studied next.184
Consider both the triangular INF , with membership functionmF (x), and the trivial INV0 in Fig.3. IN F is185
specified by the three parametersm, wL andwR. A horizontal line at heighth ∈ (0, 1] intersects INF at points186
ah and bh; moreover, it intersects trivial INV0 at pointsch and dh, wherech = dh = V0. Since the left line of187
the triangular membership functionmF (x) equalsy = [x − (m − wL)]/wL and the right line ofmF (x) equals188
y = [(m + wR) − x]/wR, it follows ah = wLh + (m − wL), moreoverbh = −wRh + (m + wR). Next, we189
analytically calculate inclusion measuresigma-joinσg(F ¹ V0) =1∫0
v(θ(ch))+v(dh)v(θ(ah∧ch))+v(bh∨dh)dh usingv(x) = px and190
θ(x) = Q− x. Integral∫
1ax+bdx = 1
a ln|ax + b|+ C0 will be useful in the following calculations.191
(1) For m + wR ≤ V0, it follows192
σg(F ¹ V0) =1∫0
Q−ch+dh
Q−ah+dhdh = −Q
1∫0
1wLh+[(m−wL)−(Q+V0)]
dh = QwL
ln (Q+V0)−m+wL
(Q+V0)−m .193
(2) For m ≤ V0 ≤ m + wR, it follows194
σg(F ¹ V0) =h0∫0
Q−ch+dh
Q−ah+bhdh+
1∫h0
Q−ch+dh
Q−ah+dhdh = −Q
h0∫0
1(wL+wR)h−(Q+wL+wR)dh−Q
1∫h0
1wLh−[Q−(m−wL)+V0]
dh =195
QwL+wR
ln Q+wL+wR
(Q+wL+wR)−(wL+wR)h0+ Q
wLln [Q−(m−wL)+V0]−wLh0
[Q−(m−wL)+V0]−wL, whereh0 = mF (V0).196
(3) For m− wL ≤ V0 ≤ m, it follows197
σg(F ¹ V0) =h0∫0
Q−ch+dh
Q−ah+bhdh+
1∫h0
Q−ch+dh
Q−ch+bhdh = −Q
h0∫0
1(wL+wR)h−(Q+wL+wR)dh−Q
1∫h0
1wRh−[Q−V0+(m+wR)]dh =198
QwL+wR
ln Q+wL+wR
(Q+wL+wR)−(wL+wR)h0+ Q
wRln [Q−V0+(m+wR)]−wRh0
[Q−V0+(m+wR)]−wR, whereh0 = mF (V0).199
(4) For V0 ≤ m− wL, it follows200
σg(F ¹ V0) =1∫0
Q−ch+dh
Q−ch+bhdh = −Q
1∫0
1wRh−[Q−V0+(m+wR)]dh = Q
wRln (m+Q−V0)+wR
m+Q−V0.201
A triangular IN’s edge corresponds to a uniform pdf as shown in Fig.4(a) as well as in Fig.4(b). Letp1(x)202
andp2(x) be the latter pdfs corresponding to INsF1 andF2, respectively. More specifically, it is203
pi(x) =
12wL
, mi − wL ≤ x ≤ mi
12wR
, mi ≤ x ≤ mi + wR
, for i ∈ {1, 2},204
wherewL and wR represent the ranges of the uniform pdf located to the left and to the right, respectively,205
of the medianmi, i ∈ {1, 2}; hence, in Fig.4(a) it iswL = r, wR = R, whereas in Fig.4(b) it iswL = R,206
wR = r. Note that themedian“m” of a pdf p(x) is defined here such thatm∫−∞
p(x)dx = 0.5 =+∞∫m
p(x)dx. Next,207
we compute the means as well as the variances of pdfsp1(x) and p2(x) corresponding to the INsF1 and F2,208
respectively.209
µ1 =+∞∫−∞
xp1(x)dx =m1∫
m1−r
x 12r dx +
m1+R∫m1
x 12Rdx = m1 + R−r
4 .210
8
µ2 =+∞∫−∞
xp2(x)dx =m2∫
m2−R
x 12Rdx +
m2+r∫m2
x 12r dx = m2 − R−r
4 .211
σ21 =
+∞∫−∞
(x− µ1)2p1(x)dx =m1∫
m1−r
(x− µ1)2 12r dx +
m1+R∫m1
(x− µ1)2 12Rdx = 5r2+5R2+6Rr
48 .212
σ22 =
+∞∫−∞
(x− µ2)2p2(x)dx =m2∫
m2−R
(x− µ2)2 12Rdx +
m2+r∫m2
(x− µ2)2 12r dx = 5r2+5R2+6Rr
48 .213
We remark thatwL = wR implies bothµ =+∞∫−∞
xp(x)dx =m+wR∫m−wL
x 1wL+wR
dx = m andσ2 = (wL+wR)2
12 as214
expected for a uniform pdf – Recall also that a uniform pdf corresponds to an isosceles triangular IN [34], [35].215
In Fig.4(c), pdfsp1(x) and p2(x) were placed such thatµ1 = µ = µ2; the corresponding INs, respectively,216
F1 and F2 are also shown in Fig.4(c). On the one hand, note that both the first- and the second- order statistics217
of pdfs p1(x) and p2(x) are identical, i.e.µ1 = µ2 and σ1 = σ2. Nevertheless, pdfsp1(x) and p2(x) differ in218
their third-order statistic, namely theirskewness. More specifically,p1(x) is skewed to the left, whereasp2(x) is219
skewed to the right. On the other hand, recall that an inclusion measure function can detect all-order statistics [39],220
[40], [41]. Hence, in Fig.4(c), an inclusion measure can discriminate between INsF1 and INF2 induced from pdfs221
p1(x) andp2(x), respectively, as demonstrated below.222
Furthermore, let us define the following two alternative conditions/specifications (S1)|mi− V0| ≤ T and (S2)223
|µi − V0| ≤ T , for a user-defined threshold valueT , whereV0 and mi, µi for i ∈ {1, 2} as well asR, r are224
shown in Fig.4. From both Fig.4(a) and Fig.4(b) it follows that exactly0.5 of the distribution does not satisfy (S1).225
Moreover, first, from Fig.4(a) it follows that0.5+(R−r)/8R of the distribution does not satisfy (S2) and, second,226
from Fig.4(b) it follows that0.5− (R− r)/8R of the distribution does not satisfy (S2). Note also that the truth of227
inequalitymi < µi (mi > µi) indicates that the corresponding pdf is skewed to the left (right).228
III. A F UZZY LATTICE REASONING (FLR) ENSEMBLE SCHEME229
Fuzzy lattice reasoning (FLR) is a term proposed originally for a concrete classification scheme [43], where230
an inclusion measure functionσ(A ¹ B) was employed, in the lattice of hyperboxes inRN , to compute a (fuzzy)231
degree of inclusion of a hyperboxA to another oneB. It was also shown that an inclusion measureσ(., .)232
supports two different modes of reasoning, namelyGeneralized Modus Ponensand Reasoning by Analogy. More233
specifically, on the one hand,Generalized Modus Ponensis supported as follows: Given both a rule “IFvariable234
V0 is E THEN propositionp” and a proposition “variable V0 is Ep” such thatEp ¹ E, where bothEp and E235
are elements in a lattice(L,¹), it reasonably follows “propositionp”. On the other hand,Reasoning by Analogy236
is supported as follows: Given both a set of rules “IFvariable V0 is Ek THEN propositionpk”, k ∈ {1, . . . , K}237
and a proposition “variable V0 is Ep” such thatEp � Ek, for k ∈ {1, . . . , K}, it follows “propositionpJ ”, where238
J.= arg max
k∈{1,...,K}{σ(Ep ¹ Ek) < 1}.239
A FLR extension to the lattice of INs has been possible according to the following rationale. We know (see in240
section II-C) that a IN can, equivalently, be represented either by a membership function or by a set of intervals.241
Therefore, since an interval is a hyperbox in spaceR1, it follows that an inclusion measure function can be extended242
from spaceR1 to the spaceF of INs by a single integral operation. Further enhancements are proposed next.243
9
A. FLR Enhancements244
Here we propose using the term FLR to denote any decision-making based on an inclusion measure function245
σ(., .). Note that advantages of using an inclusion measureσ(., .) include, first, accommodation of nontrivial246
(granular) input data, second, activation of a rule by an input outside the rule’s support (hence, asparserule-247
base becomes “sensibly” usable) and, third, a capacity to employ alternative positive valuation functions than248
v(x) = x (the latter positive valuation is exclusively employed in the literature, implicitly). We point out that249
a parametric positive valuation function may introduce tunable nonlinearities by optimal parameter estimation250
techniques; likewise, for aparametricdual isomorphic function.251
Recent work [37] has demonstrated that conventional fuzzy inference systems (FISs) [27], [30], [53], [72]252
apply “in principle” FLR, in lattice(FN ,¹), as follows.253
A FIS, typically, includesK rulesRk, k = 1, . . . K, of the following form254
Rule Rk : IF (variable V1 is Fk,1).AND. . . . .AND.(variable VN is Fk,N ) THEN propositionpk,255
where the antecedent of ruleRk is the conjunction ofN simple propositions “variableVi is Fk,i”, i = 1, . . . N ,256
moreover the consequent “propositionpk” of rule Rk is typically either a likewise proposition (e.g. in a Mamdani257
type FIS [53]) or a polynomial (e.g. in a Sugeno type FIS [72]). Our interest here focuses on rule antecedents.258
In particular, we assume that the degree of activation of a simple proposition “variable Vi is Fk,i”, i = 1, . . . N259
by another one “variable Vi is F0,i” equalsσg(F0,i ¹ Fk,i). The following examples demonstrate some technical260
application details.261
B. FLR Examples in lattice(F,¹)262
In this work we employ solely inclusion measureσg(., .) rather thanσf(., .) because only inclusion measure263
σg(., .) is non-zero for non-overlapping INs; hence, onlyσg(., .) can reason based on asparserule base.264
Example - 1265
INs F andV0 referred to, in this example, are shown in Fig.3.266
Fig.5 plots inclusion measureσg(F ¹ V0) versus the medianm of IN F from m = 0.5 to m = 9.5 using267
parameter valueswL = wR = 0.5 andV0 = 4.6; moreover, both the linear positive valuationv(x) = px and dual268
isomorphic functionθ(x) = Q − x were used with parameter valuesp = 1, Q = 10. Equality wL = wR = 0.5269
implies that triangular INF has, in particular, an isosceles triangular shape – Recall that an isosceles triangular270
IN corresponds to a uniform pdf. Since the median (m) equals the mean (µ) of a uniform pdf it follows that, for271
an isosceles triangular IN, thex-axis in both Fig.5 and Fig.6, denotesm as well asµ.272
Fig.6 plots inclusion measureσg(F ¹ V0) versus its medianm from m = 0.5 to m = 9.5 using parameter273
valueswL = wR = 0.5 andV0 = 4.6. Moreover, both positive valuationv(x) = 11+e−0.5(x−4.6) and dual isomorphic274
function θ(x) = 2(4.6)− x were employed.275
10
Notice the similarity of Fig.5 and Fig.6, where each figure was generated using a different positive valuation276
function v(x). In particular, Fig.5 was generated using alinear positive valuation, whereas Fig.6 was generated277
using asigmoidpositive valuation. In all our experiments, in the context of this work, we empirically confirmed that278
for any linear positive valuationv`(x) there is a sigmoid positive valuationvs(x), which produces an “identical”,279
for all practical purposes, inclusion measureσg(., .) function. A sigmoid positive valuation is preferable because it280
is defined over the whole setR of real numbers, therefore no truncation/normalization is necessary. In conclusion,281
unless otherwise specified, in the remaining of this work we employ sigmoid positive valuation functions.282
Example - 2283
The previous example has dealt with isosceles (triangular) INs. This example considers non-isosceles triangular284
IN shapes towards demonstrating that an inclusion measure can effectively detect higher-order statistics.285
Fig.7(a) displays inclusion measureσg(F1 ¹ V0) versus its medianm1 from m1 = 3 to m1 = 90 using IN286
F1 parameter valueswL = r = 3, wR = R = 10 andV0 = 65; Fig.7(b) shows the latter figure in the vicinity of287
its global maximum atm1 = 65. Likewise, Fig.7(c) displays inclusion measureσg(F2 ¹ V0) versus its median288
m2 from m2 = 10 to m2 = 97 using IN F2 parameter valueswL = R = 10, wR = r = 3 andV0 = 65; Fig.7(d)289
shows the latter figure in the vicinity of its global maximum atm2 = 65. Where, INsF1, F2 andV0 are shown in290
Fig.4. Finally, Fig.7(e) displays both inclusion measuresσg(F1 ¹ V0) and σg(F2 ¹ V0) versus their (identical)291
meanµ. More specifically, Fig.7(e) demonstrates thatσg(F2 ¹ V0) reaches its global maximumbeforeV0 = 65, as292
expected, because INF2 is skewed to the right; whereas,σg(F1 ¹ V0) reaches its global maximumafter V0 = 65,293
also as expected, because INF1 is skewed to the left.294
C. FLRpe: A Pairwise FLR Ensemble Scheme for Reasoning295
Based on an expert-supplied propositionp : “Variable V equalsx” the question here is to decide whether296
another propositionp0 : “Variable V equalsx0” is true or not, where bothx andx0 are INs. We responded to the297
aforementioned question by computing a (fuzzy) degree of fulfillment of implication “p → p0” by σg(x ¹ x0).298
More specifically, ifσg(x ¹ x0) ≥ T , whereT ∈ [0, 1] is user-defined, only then propositionp0 is accepted.299
Since a single expert propositionp may be prone to errors, hence it may be unreliable, we assumed an ensemble300
of N experts each one of whom supplied one propositionpk : “Variable V equalsxk ’, k ∈ {1, . . . , N}. Our basic301
assumption is that at least 2 out of theN experts are reliable. In conclusion, FLR is carried out by considering all302
different pairs of experts as shown in Algorithm 1, that is the FLRpe scheme.303
We remark that the FLRpe scheme accepts propositionp0 if and only if the corresponding implicationspk → p0,304
k ∈ {1, . . . N} of any two expertsk ∈ {i, j} are jointly accepted, in the sense that it isσg(xk ¹ x0) ≥ T for two305
different expertsk ∈ {i, j} as indicated in the mathematical expression in the last step of Algorithm 1; the latter306
(expression) derives from Proposition 2.2. In other words, propositionp0 is accepted if and only if the maximum307
(∨
) inclusion measureσg(.) of all different pairs of experts is above a user-defined thresholdT ∈ [0, 1]. Apparently,308
the FLRpe is a “collective reasoning” scheme based on an ensemble of experts.309
11
Algorithm 1 FLRpe: A Pairwise FLR Ensemble Scheme1: Consider a propositionp0 : “Variable V equalsx0” and a thresholdT ∈ [0, 1]. Furthermore, consider
N expert-supplied propositionspk : “Variable V equalsxk”, k ∈ {1, . . . , N}, wherex0, xk are INs,k ∈ {1, . . . , N}.
2: Consider one implicationrk, k ∈ {1, . . . , N} per expert as follows:Implication rk : IF pk THEN p0, symbolicallypk → p0.
3: Compute the degreeσg(xk ¹ x0) of fulfillment of each implicationrk : pk → p0, k ∈ {1, . . . , N}.4: Accept propositionp0 if and only if∨
i,j∈{1,...,N},i6=j
σ∧([xi, xj] ¹ [x0, x0]) =∨ { ∧
i,j∈{1,...,N},i6=j
{σg(xi ¹ x0), σg(xj ¹ x0)} } ≥ T
IV. A N INDUSTRIAL DISPENSINGAPPLICATION310
This section outlines an industrial application.311
A. The Industrial Problem312
Ouzois a popular Greek liquor, whose final stage production involves dispensing three different liquids, namely313
water, spirit, andyeast, to a “mixing” tank. More specifically,water is typically supplied by a local utility company,314
spirit is a commercial product whoseGs = 96% volume is pure ethanol, moreover theyeast, whoseGy volume315
(in the range40%− 80%) is pure ethanol, is prepared according to a local recipe.316
The Greek law calls for a specific percentage (Gb1) of ethanol in the final (ouzo) product, e.g.Gb
1 = 38% or317
Gb1 = 40%, etc. Furthermore, the law calls for a specific ratiopy
1 : ps1, wherepy
1 denotes the final product’s ethanol318
percentage stemming-from-yeast andps1 denotes the corresponding percentage stemming-from-commercial-spirit;319
it is py1 + ps
1 = 1. In the context of this work, we call pair(Gb1, p
y1 : ps
1) alcoholic identityof the (ouzo) product.320
Currently, the production of ouzo is largely empirical, therefore it is prone to errors as explained next.321
Typically, a skilled worker (manually) calculates the volumes of water (V w1 ), spirit (V s
1 ), and yeast (V y1 ) required322
to produce a specific volumeV b1 of ouzo ofalcoholic identity(Gb
1, py1 : ps
1). Nevertheless, when a different volume323
V b2 6= V b
1 is requested, at the absence of a skilled worker to compute the corresponding volumesV w2 , V s
2 , andV y2 ,324
then errors may occur. Another source of errors regards the manual dispensing of volumesV w1 , V s
1 , and V y1 to325
the mixing tank. Hence, thealcoholic identityof the final (ouzo) product might be outside specifications. It is of326
practical interest to keep, an automated ouzo production, within specifications.327
Work is, currently, under way towards automating the production of ouzo for a local beverage company in328
the Greek Macedonia region. Note that the problem of industrial dispensing has been treated also by other authors329
[14] using conventional modeling techniques; moreover, fuzzy regression techniques have been employed [32]. We330
applied the FLRpe scheme via a novel software platform, developed for the needs of this work as described next.331
B. A Novel Software Platform332
A novel software platform, namely XtraSP.v1 (Fig.8), was developed for the needs of this work using the333
Labview environment of the National Semiconductors Company. XtraSP.v1 operates as a user-friendly interface for334
12
controlling all the required electromechanical equipment, including four valves and one pump, via a NI USB-6501335
device. The latter (USB) is a Universal Serial Bus to digital I/O device which also measures the flow, in the range336
6 − 120 `t/min, to the mixing tank by counting pulses generated by aflowmeterusing a 32 bit long counter.337
Mounted (inside) on the upper side of the mixing tank there is anultrasonic level meter(U.L.M.) device, which338
measures the liquid level in the mixing tank with accuracy in the range3− 6 mm by transmitting short ultrasonic339
pulses to the liquid surface. In addition, there is a transparentcommunicating tube(C.T.) connected to the side of340
the mixing tank, which (tube) functions as an indicator of the liquid level (in the mixing tank) by operating on the341
principle of communicating tubes. The overall physical system architecture is shown in the upper half of Fig.8.342
In worksheet cells of XtraSP.v1 a user can specify (a) A label, e.g. for a tank, (b) An initial quantity of343
a liquid in a tank, (c) The percentage of ethanol in both the (commercial) spirit and the yeast, (d) The total344
percentage of ethanol in the undisposed ouzo, (e) The percentages of ethanol in the undisposed ouzo stemming,345
respectively, from (commercial) spirit and yeast, (f) The desired percentage of (pure) ethanol in the mixing tank,346
(g) The desired percentages of ethanol in the mixing tank stemming, respectively, from (commercial) spirit and347
yeast. Box “DECISION-SUPPORT & PARAMETERS” allows the user to specify useful rules & parameters.348
Software platform XtraSP.v1 can automatically carry out any required calculation/action on user demand.349
Furthermore, a number of safety instructions as well as warning messages can be issued. Note also that software350
platform XtraSP.v1 can operate either in a SIMULATION mode or in a real-world OPERATION mode, where the351
latter (mode) can be either MANUAL or AUTOMATIC.352
C. Implementation of the FLRpe Scheme353
An expert-based reasoning scheme, which may also accommodate uncertainty/ambiguity, is of particular interest354
in an industrial application. Furthermore, the capacity to effectively cope with an unreliable expert is a specification355
of critical importance because an unreliable expert may result in a final product outside specifications. The proposed356
FLRpe scheme appears to satisfy the aforementioned specifications, therefore it was applied as described next.357
The volume of a liquid being dispensed to the mixing tank was estimated simultaneously by three different358
“experts” including, first, a flowmeter measurement device, second, an ultrasonic level meter measurement device359
and, third, a human expert who visually consults the transparent tube connected to the side of the mixing tank. We360
employed the following (binary) decision rule.361
Rule R : IF volumev (of the liquid being dispensed) equalsV0 THEN stop dispensing,362
We assumed that the degree of truth of a RuleR equals the degree of truth of its antecedent. Hence, we “stop363
dispensing” if the antecedent propositionp0: “volume v (of the liquid being dispensed) equalsV0” is true. The364
latter (antecedent) degree of truth was calculated from the degrees of fulfillmentσg(Vi ¹ V0) of implications365
rk : IF “volume v is Vk” THEN “volume v is V0”,366
where one implicationrk, k ∈ {1, 2, 3} was supplied per expert.367
13
Therefore, the FLRpe scheme was applied as described in Algorithm 1. We point out that dispensing stops368
if and only if at least two volume IN estimates, supplied by two different experts, approximate volumeV0 in an369
inclusion measure “σg(., .) ≥ T ” sense for a user-defined thresholdT .370
V. EXPERIMENTS AND RESULTS371
We carried out comparative simulation experiments as described in this section.372
A. Disparate Data Representation and Fusion373
Recall that the FLRpe scheme here consists of an ensemble of three experts including Expert-1, that is a374
flowmeter measurement device, Expert-2, that is an ultrasonic level meter device and, Expert-3, that is a human375
expert supervisor of the industrial dispensing procedure.376
First, a dispensed (liquid) volume estimate supplied by Expert-1 was represented by a triangular IN (Fig.9)377
as follows. Even though our flowmeter device supplies a precise measurement, there is uncertainty regarding the378
dispensed volume due to both time-delays and the storage capacity of the pipes used to drive a fluid to the mixing379
tank. The latter uncertainty was modeled by two adjacent uniform pdfs, respectively, one above- and the other below-380
an obtained flowmeter measurement. For instance, let a flowmeter measurement be eitherm1 (Fig.4(a)) orm2381
(Fig.4(b)). The aforementioned two adjacent uniform pdfs are shown in Fig.4(a) as well as Fig.4(b). In conclusion,382
an estimate for a dispensed liquid volume by Expert-1 had a triangular shape as in Fig.9. The corresponding383
inclusion measure functionσg(F ¹ V0), for V0 = 65, is plotted in Fig.10 versus the medianm.384
Second, a dispensed (liquid) volume estimate supplied by Expert-2 was represented by an irregularly shaped IN385
(Fig.11) as follows. In a short sequence, we obtained a number ofN = 9 successive measurements of the liquid level386
in the mixing tank resulting in a population ofN = 9 estimates of the dispensed liquid volume. In conclusion, from387
the aforementioned population, we induced a IN (Fig.11) using algorithm CALCIN. The corresponding inclusion388
measure functionσg(F ¹ V0), for V0 = 65, is plotted in Fig.12 versus the medianm.389
Third, a dispensed (liquid) volume estimate supplied by Expert-3 was represented by a trapezoidal IN (Fig.13)390
as follows. A human supervisor of the industrial procedure, based on visual inspection of the transparent tube391
connected to the side of the mixing tank (Fig.8) as well as based on personal judgement, supplied a numeric392
estimatem of the middle of an interval[m − w, m + w] which (interval) is the core of a trapezoidal fuzzy set.393
Furthermore, both trapezoidal tailswL and wR in Fig.13 were suggested by Expert-3. Fig.14 displays a typical394
estimate for a dispensed liquid volume given by Expert-3, wherew = 1, wL = 5 andwR = 2. The corresponding395
inclusion measure functionσg(F ¹ V0), for V0 = 65, is plotted in Fig.15 versus the medianm.396
We remark that both curves in Fig.10 and Fig.15 are smooth because they have been computed analytically using397
equations in section II-E; whereas, the curve in Fig.12 is not smooth due to the irregularly shaped IN of Fig.11.398
Furthermore note that, first, the triangular IN (Fig.4) supplied by Expert-1 represents a probability distribution399
includinga priori information; in particular, the two adjacent iniform pdfs in either Fig.4(a) or Fig.4(b) representa400
14
priori information supplied by the user. Second, the irregularly shaped IN (Fig.11) supplied by Expert-2 represents401
a distribution of measurements and, third, the trapezoidal IN (Fig.14) supplied by Expert-3 represents a fuzzy set.402
Hence, each expert interprets differently the IN it supplies. In the latter sense, disparate data fussion takes place.403
B. Comparative Experimental Results and Discussion404
We carried out, comparatively, preliminary computer simulation experiments, using a standard commercial405
software package (MATLAB), as described in the following.406
First, we compared an employment of the meanµ versus the medianm of a distribution. Note that a standard407
practice in the industry is to employ the average/mean valueµ of a population of measurements instead of the408
corresponding median valuem as it was demonstrated above (see in section III-B, Example-2). However, the409
theoretical discussion above (see in the last paragraph of section II-E) has shown that an employment of inequality410
m < µ, for skewed pdfs, can increase the probability of a dispensed liquid volume “being inside the specifications”.411
In a series of Monte-Carlo computer experiments we confirmed, for both Expert-1 and Expert-2, that a combined412
employment ofm andµ results in fewer violations of the specifications. The latter is significant for our industrial413
application. Nevertheless, a conceptual problem arises regarding the employment of a medianm computed for the414
fuzzy set supplied by Expert-3 because a medianm is meaningless for a fuzzy set. However, due to the one-to-415
one correspondence between INs and pdfs [34], [35], [39], [40], it follows that for any IN amedian equivalent416
(parameter)m can be defined. Moreover, compared with the medianm of a pdf, inclusion measureσg(.) has the417
advantage that onlyσg(.) can capture higher-order data statistics; in fact,σg(.) can capture all-order data statistics418
[39], [40], [41].419
Second, we comparatively evaluated the performance of our proposed FLRpe scheme. The latter (scheme) was420
tested in a number of computer simulation experiments assuming a single unreliable expert. More specifically, we421
assumed that two experts were able to supply accurate (dispensed) liquid volume INs, whereas the third expert422
supplied a IN either at random or lagging/leading the correct volume. In other words, we used “intact” two of the423
three inclusion measuresσg(F ¹ V0) curves shown in Fig.10, Fig.12 and Fig.15, whereas we used either random424
samples of the third curve or a left/right-translated version of the third curve. In conclusion, analternative decision425
schemehas employed the average of the three inclusion measures values supplied by the three experts.426
Each one of the three inclusion measuresσg(F ¹ V0) curves shown in Fig.10, Fig.12 and Fig.15 was sampled427
at specific values of the parameterm – Note that successive parameterm samples correspond to successive time428
instances. Then, both the FLRpe and the aforementionedalternative decision schemewere applied at every (data)429
sampling instance. We confirmed, using thresholdT = 0.93, that the FLRpe scheme always accurately stops430
dispensing, whereas the alternative decision scheme may fail even at all (data) sampling instances. Note also that431
a single expert never performed better than the FLRpe scheme. Such reliable decision-making, as the FLRpe can432
provide, can be of critical importance in our industrial application due to the fact that one of the three experts may,433
occasionally, fail as it will be detailed in a future publication.434
15
VI. CONCLUSION435
Automated as well as accurate dispensing towards retaining a competitive product quality is of interest in a436
wide range of industrial applications including plastics, chemicals, dyeing, pharmaceuticals, and foods. This work437
has demonstrated a novel scheme, namelyFuzzy Lattice Reasoning pairwise ensemble, or FLRpe for short, for438
industrial dispensing based on (FLR) reasoning, which may accommodate imprecision/uncertainty/vagueness in the439
data. The FLRpe operates by considering, pairwise, all combinations of a number of expert implications based on440
the sigma-joinσg(., .) inclusion measure. Preliminary experimental results have been encouraging.441
This work has also presented a formal information fusion framework, namely the Cartesian product lattice442
(FN ,¹) of Intervals’Numbers (INs), towards an integration of disparate types of data including (intervals of)443
real numbers as well as probability/possibility distributions. Furthermore, a number of mathematical improvements444
were presented. Several illustrative examples have demonstrated practical advantages of the proposed techniques445
including the employment of granular input data as well as the sensible employment of a sparse rule base.446
Future plans include, first, a study of implicationp → q based on both inclusion measuresσf(., .) andσg(., .)447
and, second, an industrial application of the FLRpe scheme for automated ouzo production. The mathematical448
instruments presented here may also be especially useful for the design of dynamically evolving fuzzy systems [4],449
as well as for fuzzy regression analysis [8].450
APPENDIX451
This Appendix summarizes useful notions and tools regarding lattice theory [7], [35], [43], [59] using an452
improved mathematical notation [31], [37].453
A. Mathematical Background454
Given a setP , a binary relation (¹) in P is called partial order if and only if it satisfies the following455
conditions:x ¹ x (reflexivity), x ¹ y and y ¹ x ⇒ x = y (antisymmetry), and x ¹ y and y ¹ z ⇒ x ¹ z456
(transitivity) – We remark that theantisymmetrycondition may be replaced by the following equivalent condition:457
x ¹ y andx 6= y ⇒ y � x. If both x ¹ y andx 6= y then we writex ≺ y. A partially ordered set, or posetfor458
short, is a pair(P,¹), whereP is a set and¹ is a partial order relation inP . Note that, in this work, we employ459
an improved mathematical notation using, first, “curly” symbolsg, f, ¹, ≺, etc. for general poset/lattice elements460
and, second, “straight” symbols such as∨, ∧, ≤, <, etc. for real numbers, i.e. elements of the totally-ordered461
lattice (R,≤).462
A lattice is a poset(L,¹) any two of whose elementsx, y ∈ L have both agreatest lower bound, or meetfor463
short, and aleast upper bound, or join for short, denoted byx f y andx g y, respectively. Two elementsx, y ∈ L464
in a lattice(L,¹) are calledcomparable, symbolicallyx ∼ y, if and only if it is eitherx ¹ y or x  y. A lattice465
(L,¹) is calledtotally-orderedif and only if x ∼ y for any x, y ∈ L. If x � y holds for two elementsx, y ∈ L of466
a lattice(L,¹) thenx andy are calledincomparableor, equivalently,parallel, symbolicallyx||y.467
16
Given a lattice(L,¹) it is known that(L,¹∂) ≡ (L,º) is also a lattice, namelydual (lattice), where¹∂ denotes468
the dual (i.e. converse) of order relation¹. Furthermore, it is known that the Cartesian product(L1,¹)× (L2,¹),469
of two lattices(L1,¹) and (L2,¹), is a lattice with order(x1, x2) ¹ (y1, y2) ⇔ x1 ¹ y1 and x2 ¹ y2. In the470
latter Cartesian product lattice it holds both(x1, x2) f (y1, y2) = (x1 f y1, x2 f y2) and (x1, x2) g (y1, y2) =471
(x1 g y1, x2 g y2). It follows that the Cartesian product(L,º) × (L,¹) ≡ (L × L,º × ¹) is a lattice with472
order (x1, x2) ¹ (y1, y2) ⇔ x1 º y1 and x2 ¹ y2; moreover,(x1, x2) f (y1, y2) = (x1 g y1, x2 f y2) and473
(x1, x2) g (y1, y2) = (x1 f y1, x2 g y2). An element of lattice(L × L,º × ¹) will be denoted by a pair ofL474
elements within square brackets, e.g.[a, b].475
Our interest, here, is incompletelattices. Recall that a lattice(L,¹) is calledcompletewhen each of its subsets476
X has both a greatest lower bound and a least upper bound inL; hence, forX = L it follows that a complete477
lattice has both aleastand agreatestelement. In the interest of simplicity, here we use the same symbolsO and478
I to denote the least and the greatest element, respectively, in any complete lattice. Likewise, we use the same479
symbol¹ to denote the partial order relation in any (complete) lattice. Consider the following definition.480
Definition 4: Let (L,¹) be a complete lattice with least and greatest elementsO and I, respectively. An481
inclusion measurein (L,¹) is a functionσ : L× L → [0, 1], which satisfies the following conditions482
I0. σ(x,O) = 0, ∀x 6= O.483
I1. σ(x, x) = 1,∀x ∈ L.484
I2. x f y ≺ x ⇒ σ(x, y) < 1.485
I3. u ¹ w ⇒ σ(x, u) ≤ σ(x,w).486
We remark that an inclusion measureσ(x, y) can be interpreted as the fuzzy degree to whichx is less thany;487
therefore notationσ(x ¹ y) may be used instead ofσ(x, y).488
B. Useful Mathematical Instruments489
Two different inclusion measures are presented next, based on apositive valuation3 function.490
Theorem 6.1:Let functionv : L → R be a positive valuation in a complete lattice(L,¹) such thatv(O) = 0;491
then both functionssigma-meetσf(x, y) = v(xfy)v(x) andsigma-joinσg(x, y) = v(y)
v(xgy) are inclusion measures.492
Due to practical restrictions, we introduce two constraints on positive valuation functions, next. First, in order493
to satisfy condition I0 of Definition 4, our interest is in positive valuation functions such that “v(O) = 0”. Second,494
since a positive valuation functionv : L → R implies a metric (distance) functiond : L × L → R≥0 given by495
d(a, b) = v(a g b)− v(a f b), furthermore infinite distances between lattice elements are not desired, our second496
constraint is “v(I) < +∞”. Our interest, in the context of this work, focuses solely on inclusion measure functions.497
3Positive valuationin a general lattice(L,¹) is a real functionv : L× L → R that satisfies bothv(x) + v(y) = v(x f y) + v(x g y) andx ≺ y ⇒ v(x) < v(y).
17
A bijective (i.e. one-to-one)dual isomorphic4 function θ : L → L such thatx ≺ y ⇔ θ(x) Â θ(y), in a498
lattice (L,¹), can be used for extending an inclusion measure from a lattice(L,¹) to the corresponding lattice of499
intervals. Given a dual isomorphic functionθ : L → L there follow, by definition, bothθ(xf y) = θ(x)g θ(y) and500
θ(x g y) = θ(x) f θ(y). The latter equalities are handy in the proof of the following Proposition.501
Proposition 6.2: Let real functionv : L → R be apositive valuationin a lattice(L,¹); moreover, let bijective502
function θ : L → L be dual isomorphicin (L,¹), i.e. x ≺ y ⇔ θ(x) Â θ(y). Then, functionv∆ : L× L → R given503
by v∆(a, b) = v(θ(a)) + v(b) is a positive valuation in lattice(L× L,º × ¹).504
Proof505
1. First, we show thatv∆(a, b) + v∆(c, d) = v∆((a, b) f (c, d)) + v∆((a, b) g (c, d)) as follows.506
v∆(a, b) + v∆(c, d) = [v(θ(a)) + v(b)] + [v(θ(c)) + v(d)] = [v(θ(a)) + v(θ(c))] + [v(b) + v(d)] = [v(θ(a) f507
θ(c))+v(θ(a)gθ(c))]+[v(bfd)+v(bgd)] = [v(θ(agc))+v(θ(afc))]+[v(bfd)+v(bgd)] = [v(θ(agc))+508
v(bfd)]+[v(θ(afc))+v(bgd)] = v∆(agc, bfd))+v∆(afc, bgd) = v∆((a, b)f(c, d))+v∆((a, b)g(c, d)).509
2. Second, we show that(a, b) ≺ (c, d) ⇒ v∆(a, b) < v∆(c, d) as follows.510
(a, b) ≺ (c, d) ⇒ either (a  c and b ¹ d) or (a º c and b ≺ d) ⇒ either (θ(a) ≺ θ(c) and b ¹ d)511
or (θ(a) ¹ θ(c) and b ≺ d) ⇒ either (v(θ(a)) < v(θ(c)) and v(b) ≤ v(d)) or (v(θ(a)) ≤ v(θ(c)) and512
v(b) < v(d)) ⇒ v(θ(a)) + v(b) < v(θ(c)) + v(d) ⇒ v∆(a, b) < v∆(c, d).513
The latter completes the proof of Proposition 6.2.514
We remark that Proposition 6.2 has been proven, quite restrictively, for a totally-ordered lattice(L,¹) in [43].515
Acknowledgement516
This work has been supported, in part, by a project Archimedes-III contract.517
REFERENCES518
[1] A. Abraham, E. Corchado, J.M. Corchado, Hybrid learning machines – Guest Editorial, Neurocomputing 72 (13-15) (2009)519
2729–2730.520
[2] C. Alcalde, A. Burusco, R. Fuentes-Gonzalez, A constructive method for the definition of interval-valued fuzzy implication521
operators, Fuzzy Sets and Systems 153 (2) (2005) 211–227.522
[3] G.P. Amis, G.A. Carpenter, Self-supervised ARTMAP, Neural Networks 23 (2) (2010) 265–282.523
[4] P. Angelov, D. Filev, N. Kasabov, Evolving Fuzzy Systems – Guest Editorial, IEEE Transactions on Fuzzy Systems 16524
(6) (2008) 1390–1392.525
[5] H.G. Ayad, M.S. Kamel, On voting-based consensus of cluster ensembles, Pattern Recognition 43 (5) (2010) 1943–1953.526
4A function ψ : (P,¹) → (Q,¹), between posets(P,¹) and (Q,¹), is called(order) isomorphiciff both “x ¹ y ⇔ ψ(x) ¹ ψ(y)”and “ψ is ontoQ”; then, posets(P,¹) and (Q,¹) are calledisomorphic, symbolically(P,¹) ∼= (Q,¹).
18
[6] S. Binsaeid, S. Asfour, S. Cho, A. Onar, Machine ensemble approach for simultaneous detection of transient and gradual527
abnormalities in end milling using multisensor fusion, Journal of Materials Processing Technology 209 (10) (2009) 4728–528
4738.529
[7] G. Birkhoff, Lattice Theory, American Mathematical Society, 1967, Colloquium Publications 25.530
[8] A. Bisserier, R. Boukezzoula, S. Galichet, A revisited approach to linear fuzzy regression using trapezoidal fuzzy intervals,531
Information Sciences 180 (19) (2010) 3653–3673.532
[9] E. Bosse, From sensing to making sense to decision support, plenary speech delivered in the Hybrid Artificial Intelligence533
Systems (HAIS ’2010) International Conference, 23-25 June 2010, San Sebastian, Spain.534
[10] L. Breiman, Bagging predictors, Machine Learning 24 (2) (1996) 123-140.535
[11] G.A. Carpenter, S. Grossberg, N. Markuzon, J.H. Reynolds, D.B. Rosen, Fuzzy ARTMAP: a neural network architecture536
for incremental supervised learning of analog multidimensional maps, IEEE Transactions on Neural Networks 3 (5) (1992)537
698–713.538
[12] G.A. Carpenter, S. Martens, O.J. Ogas, Self-organizing information fusion and hierarchical knowledge discovery: a new539
framework using ARTMAP neural networks, Neural Networks 18 (3) (2005) 287–295.540
[13] M.C. Casey, R.I. Damper, Biologically-inspired information fusion – Guest Editorial, Information Fusion 11 (1) (2010)541
2–3.542
[14] X.B. Chen, Modeling and control of fluid dispensing processes: a state-of-the-art review, The International Journal of543
Advanced Manufacturing Technology 43 (3-4) (2009) 276–286.544
[15] E. Corchado, A. Abraham, A. de Carvalho, Hybrid intelligent algorithms and applications – Guest Editorial, Information545
Sciences 180 (14) (2010) 2633–2634.546
[16] E.M. Dos Santos, R. Sabourin, P. Maupin, Overfitting cautious selection of classifier ensembles with genetic algorithms,547
Information Fusion 10 (2) (2009) 150–162.548
[17] N.-E. El Faouzi, H. Leung, A. Kurian, Data fusion in intelligent transportation systems: progress and challenges - a survey,549
Information Fusion (in press).550
[18] M.C. Florea, A.-L. Jousselme,E. Bosse, D. Grenier, Robust combination rules for evidence theory, Information Fusion551
10 (2) (2009) 183–197.552
[19] M. Georgiopoulos, H. Fernlund, G. Bebis, G.L. Heileman, Order of search in fuzzy ART and fuzzy ARTMAP: effect of553
the choice parameter, Neural Networks 9 (9) (1996) 1541–1559.554
[20] G. Gigli, E. Bosse, G.A. Lampropoulos, An optimized architecture for classification combining data fusion and data-mining,555
Information Fusion 8 (4) (2007) 366–378.556
[21] V. Gomez-Verdejo, J. Arenas-Garcıa, A.R. Figueiras-Vidal, Committees of Adaboost ensembles with modified emphasis557
functions, Neurocomputing 73 (7-9) (2010) 1289–1292.558
[22] S.K. Goumas, I.N. Dimou, M.E. Zervakis, Combination of multiple classifiers for post-placement quality inspection of559
components: a comparative study, Information Fusion 11 (2) (2010) 149–162.560
[23] M. Grana, Lattice computing and natural computing – Guest Editorial, Neurocomputing 72 (10-12) (2009) 2065–2066.561
[24] M. Grana, I, Villaverde, J.O. Maldonado, C. Hernandez, Two lattice computing approaches for the unsupervised562
segmentation of hyperspectral images, Neurocomputing 72 (10-12) (2009) 2111–2120.563
[25] M. Grana, A.M. Savio, M. Garcıa-Sebastian, E. Fernandez, A lattice computing approach for on-line fMRI analysis, Image564
and Vision Computing 28 (7) (2010) 1155–1161.565
19
[26] A.M. Green, D.E. Angelaki, Multisensory integration: resolving sensory ambiguities to build novel representations, Current566
Opinion in Neurobiology 20 (3) (2010) 353–360.567
[27] S. Guillaume, Designing fuzzy inference systems from data: an interpretability-oriented review, IEEE Transactions on568
Fuzzy Systems 9 (3) (2001) 426–443.569
[28] A.P.M. Henriques, A.D. Doria Neto, R.F. Amaral, Classification of multispectral images in coral environments using a570
hybrid of classifier ensembles, Neurocomputing 73 (7-9) (2010) 1256–1264.571
[29] E. Hullermeier, S. Vanderlooy, Combining predictions in pairwise classification: an optimal adaptive voting strategy and572
its relation to weighted voting, Pattern Recognition 43 (1) (2010) 128–142.573
[30] J.-S.R. Jang, C.-T. Sun, Neuro-fuzzy modeling and control, Proceedings of the IEEE 83 (3) (1995) 378–406.574
[31] C. Joslyn, Order metrics for semantic knowledge systems, in: E. Corchado, M. Grana, A.M. Savio (eds.) Proc., Part II,575
5th Intl. Conf. Hybrid Artificial Intelligence Systems (HAIS ’10), San Sebastian, Spain, 23-25 June 2010, Springer, 2010,576
ser. LNAI, vol. 6077, pp. 399–409.577
[32] C.K.W. Ip, C.K. Kwong, H. Bai, Y.C. Tsim, The process modelling of epoxy dispensing for microchip encapsulation578
using fuzzy linear regression with fuzzy intervals, The International Journal of Advanced Manufacturing Technology 22579
(5-6) (2003) 417–423.580
[33] M.I. Jordan, L. Xu, Convergence results for the EM approach to mixtures of experts architectures, Neural Networks 8 (9)581
(1995) 1409–1431.582
[34] V.G. Kaburlasos, FINs: lattice theoretic tools for improving prediction of sugar production from populations of583
measurements, IEEE Transactions on Systems, Man & Cybernetics - Part B 34 (2) (2004) 1017–1030.584
[35] V.G. Kaburlasos, Towards a Unified Modeling and Knowledge-Representation Based on Lattice Theory, Springer, 2006,585
ser. Studies in Computational Intelligence 27.586
[36] V.G. Kaburlasos, Granular enhancement of fuzzy-ART/SOM neural classifiers based on lattice theory, in: V.G. Kaburlasos,587
G.X. Ritter (eds.) Computational Intelligence Based on Lattice Theory, Springer, 2007, ser. Studies in Computational588
Intelligence 67, pp. 3–23.589
[37] V.G. Kaburlasos, Granular fuzzy inference system (FIS) design by lattice computing, in: E. Corchado, M. Grana, A.M.590
Savio (eds.) Proc., Part II, 5th Intl. Conf. Hybrid Artificial Intelligence Systems (HAIS ’10), San Sebastian, Spain, 23-25591
June 2010, Springer, 2010, ser. LNAI 6077, pp. 410–417.592
[38] V.G. Kaburlasos and A. Kehagias, Novel fuzzy inference system (FIS) analysis and design based on lattice theory. part I:593
working principles, International Journal of General Systems 35 (1) (2006) 45–67.594
[39] V.G. Kaburlasos, A. Kehagias, Novel fuzzy inference system (FIS) analysis and design based on lattice theory, IEEE595
Transactions on Fuzzy Systems 15(2) (2007) 243–260.596
[40] V.G. Kaburlasos, S.E. Papadakis, Granular self-organizing map (grSOM) for structure identification, Neural Networks 19597
(5) (2006) 623–643.598
[41] V.G. Kaburlasos, S.E. Papadakis, A granular extension of the fuzzy-ARTMAP (FAM) neural classifier based on fuzzy599
lattice reasoning (FLR), Neurocomputing 72 (10-12) (2009) 2067–2078.600
[42] V.G. Kaburlasos, V. Petridis, Fuzzy lattice neurocomputing (FLN) models, Neural Networks 13 (10) (2000) 1145–1169.601
[43] V.G. Kaburlasos, I.N. Athanasiadis, P.A. Mitkas, Fuzzy lattice reasoning (FLR) classifier and its application for ambient602
ozone estimation, International Journal of Approximate Reasoning 45 (1) (2007) 152–188.603
[44] A. Kehagias, V. Petridis, V.G. Kaburlasos, P. Fragkou, A comparison of word- and sense-based text categorization using604
several classification algorithms, Journal of Intelligent Information Systems 21 (3) (2003) 227–247.605
20
[45] J. Kittler, F.M. Alkoot, Sum versus vote fusion in multiple classifier systems, IEEE Transactions on Pattern Analysis &606
Machine Intelligence 25 (1) (2003) 110–115.607
[46] R.L. Kodell, B.A. Pearce, S. Baek, H. Moon, H. Ahn, J.F. Young, J.J. Chen, A model-free ensemble method for class608
prediction with application to biomedical decision making, Artificial Intelligence in Medicine 46 (3) (2009) 267–276.609
[47] L.I. Kuncheva, Switching between selection and fusion in combining classifiers: an experiment, IEEE Transactions on610
Systems, Man & Cybernetics - Part B 32 (2) (2002) 146–156.611
[48] L.I. Kuncheva, Diversity in multiple classifier systems – Guest Editorial, Information Fusion 6 (1) (2005) 3–4.612
[49] L.I. Kuncheva, J.J. Rodrıguez, Classifier ensembles for fMRI data analysis: an experiment, Magnetic Resonance Imaging613
28 (4) (2010) 583–593.614
[50] D.-J. Lee, S. Antani, Y. Chang, K. Gledhill, L.R. Long, P. Christensen, CBIR of spine X-ray images on inter-vertebral615
disc space and shape profiles using feature ranking and voting consensus, Data & Knowledge Engineering 68 (12) (2009)616
1359–1369.617
[51] H. Li, J. Sun, Majority voting combination of multiple case-based reasoning for financial distress prediction, Expert618
Systems with Applications 36 (3, Part 1) (2009) 4363–4373.619
[52] E. Magosso, C. Cuppini, A. Serino, G. Di Pellegrino, M. Ursino, A theoretical study of multisensory integration in the620
superior colliculus by a neural network model, Neural Networks 21 (6) (2008) 817–829.621
[53] E.H. Mamdani, S. Assilian, An experiment in linguistic synthesis with a fuzzy logic controller, International Journal of622
Man-Machine Studies 7 (1) (1975) 1–13.623
[54] G. Martinez-Munoz, D. Hernandez-Lobato, A. Suarez, An analysis of ensemble pruning techniques based on ordered624
aggregation, IEEE Transactions on Pattern Analysis & Machine Intelligence 31 (2) (2009) 245–259.625
[55] M.-H. Masson, T. Denoeux, Ensemble clustering in the belief functions framework, International Journal of Approximate626
Reasoning (in press).627
[56] L. McCowan, D. Gatica-Perez, S. Bengio, G. Lathoud, M. Barnard, D. Zhang, Automatic analysis of multimodal group628
actions in meetings, IEEE Transactions on Pattern Analysis & Machine Intelligence 27 (3) (2005) 305–317.629
[57] M. Nachtegael, P. Sussner, T. Melange, E.E. Kerre, On the role of complete lattices in mathematical morphology: From630
tool to uncertainty model, Information Sciences (in press).631
[58] N.C. Oza, K. Tumer, Applications of ensemble methods – Guest Editorial, Information Fusion 9 (1) (2008) 2–3.632
[59] S.E. Papadakis, V.G. Kaburlasos, Piecewise-linear approximation of nonlinear models based on probabilistically/possibilis-633
tically interpreted intervals’ numbers (INs), Information Sciences (in press).634
[60] O. Parsons and G.A. Carpenter, ARTMAP neural networks for information fusion and data mining: map production and635
target recognition methodologies, Neural Networks 16 (7) (2003) 1075–1089.636
[61] W. Pedrycz, A. Skowron, V. Kreinovich (eds.), Handbook of Granular Computing, John Wiley & Sons, 2008.637
[62] V. Petridis, V.G. Kaburlasos, FINkNN: a fuzzy interval number k-nearest neighbor classifier for prediction of sugar638
production from populations of samples, Journal of Machine Learning Research 4 (2003) 17–37.639
[63] N. Poh, T. Bourlai, J. Kittler, L. Allano, F. Alonso-Fernandez, O. Ambekar, J. Baker, B. Dorizzi, O. Fatukasi, J. Fierrez,640
H. Ganster, J. Ortega-Garcia, D. Maurer, A.A. Salah, T. Scheidat, C. Vielhauer, Benchmarking quality-dependent and cost-641
sensitive score-level multimodal biometric fusion algorithms, IEEE Transactions on Information Forensics and Security 4642
(4) (2009) 849–866.643
[64] R. Polikar, Ensemble based systems in decision making, IEEE Circuits & Systems Magazine 6 (3) (2006) 21–45.644
21
[65] M. Re, G. Valentini, Integration of heterogeneous data sources for gene function prediction using decision templates and645
ensembles of learning machines, Neurocomputing 73 (7-9) (2010) 1533–1537.646
[66] C. Ronse, Why mathematical morphology needs complete lattices, Signal Processing 21 (2) (1990) 129–154.647
[67] K. Sambhoos, C. Bowman, J. Llinas, A case study with design of experiments: Performance evaluation methodology for648
Level 1 distributed data fusion processes, Information Fusion (in press).649
[68] R.E. Schapire, The strength of weak learnability, Machine Learning 5 (2) (1990) 197-227.650
[69] M. Shi, A. Bermak, S.B. Belhouari, P.C.H. Chan, Gas identification based on committee machine for microelectronic gas651
sensor, IEEE Transactions on Instrumentation and Measurement 55 (5) (2006) 1786–1793.652
[70] C. Silva, U. Lotric, B. Ribeiro, A. Dobnikar, Distributed text classification with an ensemble kernel-based learning approach,653
IEEE Transactions on Systems, Man & Cybernetics - Part C 40 (3) (2010) 287–297.654
[71] P. Sussner, E.L. Esmi, Morphological perceptrons with competitive learning: Lattice-theoretical framework and constructive655
learning algorithm, Information Sciences (in press).656
[72] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to modeling and control, IEEE Transactions657
on Systems, Man & Cybernetics 15 (1) (1985) 116–132.658
[73] H.-X. Tian, Z.-Z. Mao, An ensemble ELM based on modified AdaBoost.RT algorithm for predicting the temperature of659
molten steel in ladle furnace, IEEE Transactions on Automation Science & Engineering 7 (1) (2010) 73–80.660
[74] K. Uehara, M. Fujise, Fuzzy inference based on families ofα-level sets, IEEE Transactions on Fuzzy Systems 1 (2)661
(1993) 111–124.662
[75] K. Uehara, K. Hirota, Parallel and multistage fuzzy inference based on families ofα-level sets, Information Sciences 106663
(1-2) (1998) 159–195.664
[76] K. Uehara, T. Koyama, K. Hirota, Inference based onα-cut and generalized mean with fuzzy tautological rules, Journal665
of Advanced Computational Intelligence and Intelligent Informatics 14 (1) (2010) 76–88.666
[77] K. Uehara, T. Koyama, K. Hirota, Suppression effect ofα-cut based inference on consequence deviations, Journal of667
Advanced Computational Intelligence and Intelligent Informatics 14 (3) (2010) 256–271.668
[78] P.P. Wang, Mathematics of uncertainty – Guest Editorial, Information Sciences, 177 (23) (2007) 5141–5142.669
[79] X. Wang, S. Wang, D. Bi, Compacted probabilistic visual target classification with committee decision in wireless670
multimedia sensor networks, IEEE Sensors Journal 9 (4) (2009) 346–353.671
[80] M.A. Wiering, H. van Hasselt, Ensemble algorithms in reinforcement learning, IEEE Transactions on Systems, Man &672
Cybernetics - Part B 38 (4) (2008) 930–936.673
[81] M. Wozniak, Proposition of common classifier construction for pattern recognition with context task, Knowledge-Based674
Systems 19 (8) (2006) 617–624.675
[82] L.A. Zadeh, The concept of a linguistic variable and its application to approximate reasoning - I, Information Sciences 8676
(3) (1975) 199–249.677
[83] L.A. Zadeh, From computing with numbers to computing with words – from manipulation of measurements to manipulation678
of perceptions, IEEE Transactions on Circuits and Systems – I: Fundamental Theory and Applications 45 (1) (1999) 105–679
119.680
[84] P. Zhong, R. Wang, A multiple conditional random fields ensemble model for urban area detection in remote sensing681
optical images, IEEE Transactions on Geoscience and Remote Sensing 45 (12) (2007) 3978–3988.682
[85] X. Zhu, P. Zhang, X. Lin, Y. Shi, Active learning from stream data using optimal weight classifier ensemble, IEEE683
Transactions on Systems, Man & Cybernetics - Part B (in press).684
22
1.3 1.4 1.484 1.6 1.70
1
a data samples population
1.3 1.4 1.484 1.6 1.70
0.5
1
PDF: integral of the finest histogram
(a) (c)
1.3 1.4 1.484 1.6 1.70
0.1
0.2
0.3a histogram (of frequencies)
1.3 1.4 1.484 1.6 1.70
1
2*PDF
2*(1-PDF)
the corresponding IN
(b) (d)
Fig. 1. Calculation of a IN from a population of data samples. (a) The data samples with medianm = 1.484. (b) A histogram of thedata. (c) The corresponding cumulative distribution function (PDF). (d) Computation of a IN from the corresponding PDF; that is, algorithmCALCIN.
23
1.3 1.4 1.484 1.6 1.70
1F
x
mF(x
)
(a)
1.3 1.4 1.484 1.6 1.70
1F
F(x)
h
(b)
Fig. 2. The two different representations of a IN F from Fig.1(d). (a) The membership-function-representationmF (x). (b) The interval-representation forL = 32 different levels spaced evenly over the interval(0, 1].
mF
R
y
m-wL m+wRm
h
V0
ch = dhx
F1
ah bh
Fig. 3. Two INs including a triangular INF with membership functionmF , specified by the three numbersm− wL, m, m + wR, and atrivial IN V0. A horizontal line at heighth ∈ (0, 1] intersects INF at pointsah andbh, moreover it intersects trivial INV0 at ch = dh = V0.
24
V0
y
m1-r m1+Rm1x
F1
1
1
1
2r
1
2R
(a)
V0
y
m2-R m2+rm2x
F2
1
2
1
2r
1
2R
(b)
V0
y
F2
m2m1x
F1
1
(c)
Fig. 4. Triangular INsF1, F2 and trivial IN V0. (a) IN F1 corresponds to a piecewise-uniformp1(x) pdf such thatp1(x) = 12r
form1 − r ≤ x ≤ m1, whereasp1(x) = 1
2Rfor m1 ≤ x ≤ m1 + R. (b) IN F2 corresponds to a piecewise-uniformp2(x) pdf such that
p2(x) = 12R
for m2 −R ≤ x ≤ m2, whereasp2(x) = 12r
for m2 ≤ x ≤ m2 + r. (c) INs F1 andF2 were placed so as the correspondingpdfs p1(x) andp2(x), respectively, have identical means, i.e.µ1 = µ = µ2. Note that the standard deviations ofp1(x) andp2(x) are alsoidentical, i.e.σ1 = σ2.
25
1 2 3 4 4.6 5 6 7 8 9 100.5
0.6
0.7
0.8
0.9
1
h
m
(a)
3 4 4.6 5 60.8
0.9
1
h
m
(b)
Fig. 5. (a) Inclusion measureσg(F ¹ V0) is plotted versus its medianm, where INsF andV0 are shown in Fig.3, using parameter valueswL = wR = 0.5 and V0 = 4.6; moreover, both the linear positive valuationv(x) = x and the dual isomorphic functionθ(x) = 10 − xwere used. (b) The above figure is shown in the vicinity of its global maximum atm = 4.6.
26
0 1 2 3 4 4.6 5 6 7 8 9 100.5
0.6
0.7
0.8
0.9
1
h
m
(a)
3 4 4.6 5 60.8
0.9
1
h
m
(b)
Fig. 6. (a) Inclusion measureσg(F ¹ V0) is plotted versus its medianm, where INsF andV0 are shown in Fig.3 using parameter valueswL = wR = 0.5 and V0 = 4.6; moreover, both the sigmoid positive valuationv(x) = 1
1+e−0.5(x−4.6) and the dual isomorphic functionθ(x) = 2(4.6)− x were used. (b) The above figure is shown in the vicinity of its global maximum atm = 4.6.
27
0 10 20 30 40 50 60 65 70 80 90 1000.5
0.6
0.7
0.8
0.9
1
h
m1
0 10 20 30 40 50 60 65 70 80 90 1000.5
0.6
0.7
0.8
0.9
1
h
m2
(a) (c)
60 65 700.9
0.91
0.92
0.93
0.94
0.95
h
m1
60 65 700.9
0.91
0.92
0.93
0.94
0.95h
m2
(b) (d)
60 65 700.9
0.91
0.92
0.93
0.94
0.95
h
(e)
Fig. 7. INs F1, F2 and V0 are shown in Fig.4 withr = 3 and R = 10, moreover trivial INV0 is located at65. (a) Inclusion measureσg(F1 ¹ V0) is plotted versus its medianm1. (b) The latter figure is shown in the vicinity of its global maximum atm1 = 65. (c) Inclusionmeasureσg(F2 ¹ V0) is plotted versus its medianm2. (d) The latter figure is shown in the vicinity of its global maximum atm2 = 65. (e)Inclusion measuresσg(F1 ¹ V0) andσg(F2 ¹ V0) are shown, comparatively, in the vicinity of their global maximum versus their identicalmeanµ.
28
Fig. 8. A fully functional software platform, namely XtraSP.v1, has been developed, in the context of this work, towards an industrialproduction ofouzo (alcoholic) beverage by automating the corresponding liquid dispensing application. Cell label “U.L.M.” stands forUltrasonic Level Meter, moreover cell label “C.T.” stands for Communicating Tube.
29
35 40 45 50 55 60 650
0.2
0.4
0.6
0.8
1
m [lt]
(a)
35 40 45 50 55 60 650
0.2
0.4
0.6
0.8
1
m [lt]
(b)
Fig. 9. Expert-1, that is a flowmeter measurement device, supplied a triangular IN estimate of a dispensed volume as detailed in the text.(a) The membership-function-representation of a dispensed volume estimate. (b) The corresponding interval-representation.
0 10 20 30 40 50 60 65 70 80 90 1000.5
0.6
0.7
0.8
0.9
1
h
m [lt]
(a)
60 65 70 75
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
h
m [lt]
(b)
Fig. 10. (a) Inclusion measureσg(F ¹ V0) is plotted versus its medianm, where INF is shown in Fig.9, moreoverV0 = 65. (b) Theabove figure is shown in the vicinity of its global maximum atm = 65.
30
35 40 45 50 55 60 650
0.2
0.4
0.6
0.8
1
m [lt]
(a)
35 40 45 50 55 60 650
0.2
0.4
0.6
0.8
1
m [lt]
(b)
Fig. 11. Expert-2, that is a ultrasonic level meter measurement (U.L.M.) device, supplied a population of measurements resulting in a INof irregular shape as an estimate of a dispensed volume as detailed in the text. (a) The membership-function-representation of a dispensedvolume estimate. (b) The corresponding interval-representation.
0 10 20 30 40 50 60 65 70 80 90 1000.5
0.6
0.7
0.8
0.9
1
h
m
(a)
60 65 70 75
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
h
m
(b)
Fig. 12. (a) Inclusion measureσg(F ¹ V0) is plotted versus its medianm, where INF is shown in Fig.11, moreoverV0 = 65. (b) Theabove figure is shown in the vicinity of its global maximum atm = 65.
31
R
y
m-w-wL m+w+wRm
h
V0
ch=dhx
F1
ah bhm+wm-w
Fig. 13. Two INs including a trapezoidal INF, specified by the four numbersm− w − wL, m− w, m + w, m + w + wR (note thatmis the average of numbersm− w andm + w), and a trivial INV0. A horizontal line at heighth ∈ (0, 1] intersects INF at pointsah andbh, moreover it intersects trivial INV0 at pointch = dh = V0.
35 40 45 50 55 60 650
0.2
0.4
0.6
0.8
1
m [lt]
(a)
35 40 45 50 55 60 650
0.2
0.4
0.6
0.8
1
m [lt]
(b)
Fig. 14. Expert-3, that is a human expert, supplied a trapezoidal IN estimate of a dispensed volume as detailed in the text. (a) Themembership-function-representation of a dispensed volume estimate. (b) The corresponding interval-representation.
32
0 10 20 30 40 50 60 65 70 80 90 1000.5
0.6
0.7
0.8
0.9
1
h
m [lt]
(a)
60 65 70 75
0.9
0.91
0.92
0.93
0.94
0.95
0.96
0.97
0.98
0.99
1
h
m [lt]
(b)
Fig. 15. (a) Inclusion measureσg(F ¹ V0) is plotted versus its corresponding median parameterm, where IN F is shown in Fig.14,moreoverV0 = 65. (b) The above figure is shown in the vicinity of its global maximum atm = 65.