Abstract-In this paper, we present a novel technique to

generate zSlices based general type-2 fuzzy sets using a series of

interval type-2 fuzzy sets based around the notion of
"agreement" of interval type-2 fuzzy sets. We provide details
on how this approach can be applied for a series of readily
available interval type-2 fuzzy sets as well as how the proposed
approach can be employed to generate zSlices based general
type-2 fuzzy sets which are continually updated as new interval
type-2 fuzzy sets become available over time. We also describe
the proposed approach in the context of Ambient Intelligent
Environments (AIEs) which illustrate the benefits of a
continuously updated general type-2 membership function and
its potential advantage over interval type-2 fuzzy logic based
approaches. Subsequently, we demonstrate the approach based
on triangular interval type-2 fuzzy sets and we highlight the
remaining complexities and complications in terms of the
implementation of the proposed technique which stem from the
potential for the creation of non-convex fuzzy sets and propose
solutions for these problems.
I. INTRODUCTION
YPE-2 Fuzzy Logic systems (FLSs) based on type-2
Iuzzy logic sets were initially described by Zadeh in
1975 |1|. The recent years have seen a persistent rise in
interest in both type-2 FLSs applications and theoretical
research. In particular, the Iocus has been on interval type-2
FLSs which oIIer a signiIicant reduction in both complexity
and computational requirements when compared to general
type-2 FLSs. This reduction in complexity is achieved by
Iixing the secondary Membership Function (MF) oI interval
type-2 Iuzzy sets to 1 Ior all points belonging to a set
(primary memberships) which in turn Iacilitates the
computation and operation oI the type-2 FLS. Nevertheless,
this simpliIication eIIectively comes at the cost oI losing the
additional degree oI Ireedom provided by the secondary
membership in general type-2 Iuzzy sets. In other words, as
all primary memberships are associated with the same
secondary membership, it is not possible to model speciIic
primary memberships as being less uncertain than others. In
Iuzzy logic terms, in interval type-2 Iuzzy sets, the
secondary membership is modelled as an interval type-1
Iuzzy set while in general type-2 Iuzzy sets the secondary
membership is modelled as a (standard) type-1 set.

This work was supported in part by the European Commission under
Grant FP7-ICT-1-8.2/216837 entitled 'Adaptive and Trusted Ambient
Ecologies (ATRACO).
C. Wagner and H. Hagras are with the Computational Intelligence
Centre, School oI Computer Science and Electronic Engineering, University
oI Essex, Colchester CO4 3SQ, U.K. (e-mail: chwagnessex.ac.uk).
In recent years, increases in computational power and the
promise oI better uncertainty modelling (and thus -
handling) oI general type-2 Iuzzy logic have resulted in a
strong surge in the research into general type-2 Iuzzy logic
systems with encouraging results in general type-2 Iuzzy
logic set representation and operations |2|, |3|, |4|, |5|, |6|,
|7|, |8| as well as initial applications |9|, |10|.
One oI the challenges Iacing general type-2 FLSs is to
determine how best to employ the better uncertainty
modelling capabilities oIIered by general type-2 Iuzzy sets
as part oI real world applications and to evaluate their real
world perIormance in comparison to type-1, interval type-2
and other non-Iuzzy logic based systems.
In order to achieve good perIormance, general type-2
FLSs (as all FLSs) rely on an appropriately speciIied rule
base and most importantly, on accurately speciIied Iuzzy
sets. As the individual Iuzzy sets should encode most oI the
inIormation about the system to be modelled, their correct
speciIication can be highly complex. This problem is
exasperated in general type-2 Iuzzy logic sets when
compared to type-1 and interval type-2 Iuzzy sets as it is not
only the primary membership (either crisp or as an interval)
that needs to be determined, but also the secondary
membership that needs to be deIined.
In previous publications, we have shown how inIormation
about individual devices such as sensors or even about
linguistic labels can be employed to construct general type-2
Iuzzy sets |11|. In this paper, we will demonstrate a
technique which allows the gathering and combination oI
inIormation Irom multiple interval type-2 Iuzzy sets in order
to generate useIul general type-2 Iuzzy sets. A particularly
interesting application oI this technique in cases where
individual interval type-2 Iuzzy sets are generated at
consecutive time intervals is also discussed and its
useIulness and application in the Iield oI ambient
intelligence is presented.
Section (II) provides an overview oI some Iundamentals
oI zSlices and zSlices based general type-2 Iuzzy sets as
well as background inIormation on our work in ambient
intelligence, detailing how and why the proposed approach
can be applied there. Section (III) describes the process oI
generating and adapting the zSlices based general type-2
Iuzzy sets Irom a series oI interval type-2 Iuzzy sets. A
worked example is presented in Section (IV) while Section
(V) provides some conclusions and describes Iuture work.
An Approach for the Generation and Adaptation of zSlices based
General Type-2 Fuzzy Sets from Interval Type-2 Fuzzy Sets to
Model Agreement with Application to Intelligent Environments
Christian Wagner, Member, IEEE and Hani Hagras, Senior Member, IEEE
T
978-1-4244-8126-2/10/$26.00 2010 IEEE

II. BACKGROUND AND FUNDAMENTALS
As part oI the method described in this paper, a series oI
interval type-2 Iuzzy sets is employed to generate a zSlices
based general type-2 Iuzzy set which can subsequently be
updated "on-the-Ily" by introducing new interval type-2
Iuzzy sets. The concept and theory oI zSlices and zSlices
based general type-2 FLSs have been presented in |4|, |5|
and |10| and as such we are only providing the main
equations in Section (II.A) Ior reIerence. While the
generation oI the initial interval type-2 Iuzzy sets is not part
oI the Iocus oI this paper, in Section (II.B), we are brieIly
reviewing a method Ior generating interval type-2
membership Iunctions Irom data in the context oI Ambient
Intelligent Environments (AIEs).
A. zSlices and zSlices Based General Type-2 Fuzzy Sets.
As part oI this paper, we will be relying on the zSlices
based representation oI general type-2 Iuzzy sets described
in detail in |4| and |5|. To conserve space we are solely
repeating the main equations oI zSlices based general type-2
Iuzzy sets here, more details can be Iound in |4|, |5| and
|10|.
A general type-2 Iuzzy set can be represented as a total
number oI I zSlices (in a discrete universe oI discourse):

=1
(1)
where zSlice

is deIined as:

/(,

]
(2)
and i is the identiIier oI the zSlice where 0 . The
larger the number oI zSlices, the more accurate the
representation oI the original general type-2 set.
An example oI a zSlices based general type-2 Iuzzy set is
provided in Fig. 1.

(a) (b)
Fig. 1 (a) A sample zSlices based general type-2 Iuzzy set with three zSlices
(I=3). Note how each zSlice is associated with its respective zLevel i and
each zLevel is at a Iixed third dimension (secondary membership)

= /
as shown in |4| and |5|. i.e. in this case
1
=
1
3
,
2
=
2
3
and
3
= 1. (b) The
same set shown aIter a rotation in 3D space to allow a better view oI the
third dimension.
It is important to note that the only diIIerence between an
interval type-2 Iuzzy set and a zSlice is that the secondary
membership (third dimension) oI an interval type-2 Iuzzy set
is always equal to 1 while Ior zSlices it can be speciIied as
any constant

[0,1].
B. Generation of interval type-2 fuzzy sets in the context of
Ambient Intelligent Environments.
As part oI this paper, we are demonstrating the technique
oI how interval type-2 Iuzzy sets can be employed to
consecutively construct zSlices based general type-2 Iuzzy
sets using an example Irom our work in ambient
intelligence which we Ieel both Iacilitates the presentation oI
the underlying concepts and shows the real world
applicability oI the approach. In order to illustrate the
context oI this example, we are providing a brieI overview
oI the problem here. In ambient intelligence applications and
AIEs such as intelligent homes, one oI the potential
applications oI computational intelligence techniques such
as Iuzzy logic is the support oI a user in his/her everyday
environment such as his/her home through the learning oI
his/her preIerences. For example, an intelligent system
installed in a home can learn the inhabitant's preIerences in
terms oI indoor lighting, i.e. when, in what situation, at what
light level (outdoors, indoors) he/she would like which
speciIic lamps or lights in his/her home to be turned ON and
to which speciIic level oI brightness. Fig. 2 shows the iSpace
which is an AIE testbed based at the University oI Essex,
where the work presented in this paper is being developed
and deployed. The iSpace looks like any normal apartment
but it is equipped with a large number and variety oI sensors,
actuators, embedded devices and networks.

Fig. 2 The iSpace, an AIE testbed at the University oI Essex, UK
In |12|, Doctor et al. have shown how the interaction oI
the user with his/her environment (e.g. the lamps) can
transparently be monitored to create sets oI inIormation
which encode the state oI the environment and the action
which the user has taken in this speciIic state. For example,
when the user turns the kitchen lights ON, the system can
record the current values recorded by any light sensors in
range (outdoors and indoors) and create a record which
relates these states to the light level which the speciIic lamp
was set to by the user while turning it ON. Over time, many
records oI user interactions with his/her environment can be
recorded and this data can be evaluated to retrieve
inIormation about the user's preIerences.



In |12|, Doctor et al. have shown how clustering such as
Fuzzy C-Means (FCM) |13| can Iurther be employed to
extract type-1 Iuzzy sets Irom the collected data in order to
model the individual devices involved (e.g. light sensors,
lamps, etc.) and as such provide the basis Ior a subsequently
generated Iuzzy logic rule base which aims to reproduce the
user's preIerences and as in our case, to control the lights etc.
on his/her behalI.
As the behaviour oI individual devices changes (through
diIIerences in their use, wear & tear, obstruction, etc.) the
perIormance oI the type-1 Iuzzy system easily deteriorates.
This is a common occurrence with type-1 Iuzzy systems
which tend to require Irequent retuning in uncertain and
changing operating conditions unavoidable in real world
applications |14|.
In order to improve this shortcoming, in |15|, Hagras et al.
have shown how to generate interval type-2 Iuzzy sets which
provide a more resilient perIormance with the Iaced
uncertainties. Nevertheless, the main disadvantage with this
approach is that the Footprint OI Uncertainty (FOU) oI the
interval type-2 Iuzzy sets tends to grow very large as outliers
in the data (resulting in type-1 Iuzzy sets) are subsequently
incorporated into the interval type-2 Iuzzy sets. Overly large
FOUs in turn promote a slow and inexact system response
|16|.
In our case, we choose to generate interval type-2 Iuzzy
membership Iunctions Irom the initial type-1 membership
Iunctions by associating them with a FOU reIlecting the
uncertainty about the device. In terms oI light sensors Ior
example, we are employing the device precision inIormation
supplied by the manuIacturer Ior the speciIic input ranges.
While this interval type-2 Iuzzy logic approach does not
suIIer Irom interval type-2 Iuzzy sets with continuously
growing FOUs, it still, like the approach by Hagras et al.
Iails to deal with outliers and sparsely occurring conditions,
especially when considered over extended periods oI time -
where outliers could be identiIied.
In this paper, we are discussing the construction oI zSlices
based general type-2 Iuzzy sets Irom existing interval type-2
Iuzzy sets (generated Ior example once a day) which allows
the continuous updating oI the employed Iuzzy sets while
modelling "agreement" oI the individual interval type-2 sets
over time. We reIer to "agreement" as areas where individual
interval type-2 Iuzzy sets or more speciIically, their FOUs
overlap. This allows Ior the certainty that can be derived
Irom the agreement, maniIested by reoccurring similar
interval type-2 Iuzzy sets, to be employed to generate
general type-2 Iuzzy sets which can model this certainty,
while, at the same time still reIlecting the inIormation
provided by sparsely occurring interval type-2 Iuzzy sets
which can be considered as relatively uncertain. The
additional degree oI Ireedom oI general type-2 Iuzzy sets
allows the creation oI systems which can be continuously
updated without compromising the knowledge generated in
previous instances oI the system. This whole process relies
on inIormation samples which are initially expressed in the
Iorm oI interval type-2 sets being evaluated and combined
with other interval type-2 sets (created Ior example at
diIIerent times, but modelling the same device/variable) to
provide a better model based on a general type-2 Iuzzy set
and thus providing better perIormance over time.
In the Iollowing Section, we will describe the details oI
the process oI generating zSlices based general type-2 Iuzzy
sets Irom interval type-2 Iuzzy sets.
III. GENERATION OF ZSLICES BASED GENERAL TYPE-2
FUZZY SETS FROM MULTIPLE INTERVAL TYPE-2 FUZZY SETS.
The process oI generating zSlices based general type-2
Iuzzy sets described in this paper is based on interval type-2
Iuzzy sets which are incorporated to generate the Iinal
general type-2 set. Hence, it is possible to generate a zSlices
based general type-2 Iuzzy set based on a single interval
type-2 Iuzzy set and to update it continuously across
minutes, hours, days or even longer periods with Iurther
interval type-2 Iuzzy sets (as they become available e.g.
through additional data) or to generate a Iinal general type-2
Iuzzy set instantaneously based on a set oI existing interval
type-2 Iuzzy sets.
While both approaches are very similar, the continuous
updating oI general type-2 MFs is slightly more complex
while oIIering the potential to actively adjust the uncertainty
represented in the Iuzzy set over time to reIlect the changing
conditions oI the device (e.g. inIormation which is not
applicable (as Ior example the characteristics oI a device
have changed) can be dynamically removed over time). Thus
we are addressing the instantaneous creation oI Iuzzy sets
Iirst in Section (III.A) while presenting the details oI the
process oI continuous creation/updating oI Iuzzy sets in
Section (III.B).
A. Instantaneous Creation of zSlices based General Type-2
Fuzzy Sets based on multiple Interval Type-2 Fuzzy Sets.
The creation oI a zSlices based general type-2 Iuzzy set
Irom multiple interval type-2 Iuzzy sets is based on the
Iollowing basic principle:
Areas oI membership (i.e. FOUs) covered by many
interval type-2 membership Iunctions are less uncertain than
areas covered by Iewer interval type-2 membership
Iunction(s). In other words, iI a certain crisp input is
repeatedly associated with the same primary membership (or
part oI this primary membership is repeatedly the same),
then this membership (or part oI it) can be considered more
certain than memberships (or parts) which are not or
inIrequently covered by the same membership. The more
MFs overlap at a given membership value (or area), the
more certain we are about the association oI the given crisp
input to this speciIic membership value (or area). Thus,
overlapping areas will be modelled with a higher level oI
certainty, i.e. third dimension which in terms oI zSlices is
reIlected in a higher zLevel. Note that in general type-2
Iuzzy sets, increasing certainty about primary memberships
is expressed through the third dimension where 0 reIlects
complete uncertainty and 1 reIlects certainty.
While the concept is straightIorward, its implementation
can prove very challenging depending on the type oI MF
employed (triangular, gaussian, data-based piece-wise
linear,...). Furthermore, complicating Iactors such as the



common requirement Ior convexity in terms oI the generated
zSlices require Ior the designer oI the system to make a
number oI design-dependent decisions Ior resulting non-
convex Iuzzy sets (and their convex approximation). Fig. 3
illustrates the potential complexity oI overlapping FOUs
using three normal triangular interval type-2 Iuzzy sets the
parameters oI which are provided in TABLE I.
From Fig. 3 (b), (c) and (d) and Fig. 4, it is clear that the
individually identiIied overlapping areas cannot be directly
employed as a basis Ior a FOU oI a zSlice as the resulting
zSlice would be non-convex. A variety oI techniques such
as interpolating between "gaps" in the membership Iunction
could be employed to assure the convexity oI the resulting
sets - naturally at the cost oI modelling accuracy. It would be
desirable to employ non-convex sets but the theory Ior
dealing with non-convex Iuzzy sets, in particular non-
convex general type-2 Iuzzy sets remains currently
unaddressed.

(a)

(b)

(c)

(d)
Fig. 3. (a) Three interval type-2 Iuzzy sets: IT20, IT21, IT22. (b) The
same sets with a combination oI their FOUs shaded. (c) The same sets with
shaded areas where at least two FOUs overlap. (d) The same sets with
shaded areas where at least three FOUs overlap.


(a) (b)
Fig. 4. 3D visualisation oI a general type-2 Iuzzy set extracted Irom the
overlap shown shaded in Fig. 3 (d). (a) View oI the resulting set Irom the
Iront/side. (b) View oI the same set Irom the rear/below, Iacilitating the
view oI the third dimension and the area oI the set which results in it being
non-convex. Note: A 3D version oI the Figure is available online here.
However, it is worth noting how the size oI the
overlapping FOU (shaded areas in Fig. 3 (b), (c) and (d))
decreases as the number oI Iuzzy sets which are required to
overlap increases. This result is intuitive and can be
interpreted as "agreement" between the individual interval
type-2 membership Iunctions. Naturally, areas where more
interval type-2 Iuzzy sets "agree" are smaller than those
where only Iew "agree".
TABLE I
Parameters Ior Interval Type-2 Fuzzy Sets
Start Centre End
UMF LMF UMF LMF UMF LMF
IT2_0 1 2 4 4 7 6
IT2_1 0 3 4 4 7 5
IT2_3 1 3 4.5 4.5 6.5 5

Incidentally, because oI the complex nature oI the
overlapping FOUs, there is no unique conversion method Ior
converting multiple interval type-2 Iuzzy sets into zSlices
based general type-2 Iuzzy sets, but a variety oI diIIerent
conversions based on the concept oI overlapping (i.e.
"agreeing") FOUs presented in this paper are possible.
We proceed below by describing the generic steps oI the
conversion process Ior converting a series oI interval type-2
Iuzzy sets into a zSlices based general type-2 Iuzzy set. In
Section (IV), we are demonstrating a sample implementation
oI the process using the common triangular interval type-2



Iuzzy sets.
The overall process oI conversion can be separated into a
series oI generic steps:
1. Set the number oI zLevels I required. This number
can either be equal or smaller than the number oI
interval type-2 sets available. A smaller number oI
zLevels reduces the computational complexity oI
the zSlices based general type-2 Iuzzy system at the
cost oI uncertainty modelling precision |4|, |5|. In
this case multiple interval type-2 Iuzzy sets are
combined until a number oI interval type-2 sets
equal to the number oI zLevels I is achieved.
2. For each zLevel i, where 1 , determine any
areas

within the primary memberships oI the

individual interval type-2 Iuzzy sets which are
covered by at least the FOU oI i interval type-2
membership Iunctions (as shown in Fig. 3
(b),(c),(d)).
3. For each zLevel i, generate a valid zSlice

which
approximates the area

as closely as possible.
(For Iurther details on the properties oI zSlices,
please reIer to |4|, |5|.) A variety oI methods can be
applied here, we are demonstrating a technique Ior
triangular membership Iunctions in Section IV.
4. The combination oI the I zSlices gives rise to the
zSlices based general type-2 Iuzzy set

. An
example oI a zSlices with 3 zSlices is provided in
Fig. 1.
Steps 1-4 describe the basic process oI generating zSlices
based general type-2 sets Irom a series oI available interval
type-2 Iuzzy sets. This process allows the modelling oI
"agreement" between a series oI interval type-2 Iuzzy sets
while maintaining the inIormation encoded in all interval
type-2 sets (including outliers). This is achieved by
employing the additional degree oI Ireedom provided by the
third dimension in general type-2 Iuzzy sets.
In the Iollowing Section we will describe how this
approach can be extended to allow the continuous updating
oI a zSlices based general type-2 set using sequentially
available interval type-2 sets.
B. Continuous Creation and Adaptation of zSlices based
General Type-2 Fuzzy Sets based on sequentially available
Interval Type-2 Fuzzy Sets.
In many applications such as in AIEs as described in
Section (II.B), inIormation about the physical entities or
devices modelled using Iuzzy sets (e.g. lights, light
sensors,...) becomes available in a continuous Iashion, i.e.
over hours, days, weeks, etc. As time progresses, some
inIormation can become outdated and in terms oI Iuzzy sets,
this means that the properties oI the Iuzzy sets based on this
inIormation should be updated accordingly. In the same
context however, some inIormation is strengthened in terms
oI its certainty through reappearing Irequently which should
also be reIlected in the resulting Iuzzy sets.
In the context described in Section (II.B), where an
interval type-2 Iuzzy set is created at the end oI each day to
model Ior example a light sensor, the approach described in
Section (III.A) can be applied to combine the individual
interval type-2 Iuzzy sets generated across Ior example a Iull
week (7 days). As was described, this will result in a zSlices
based general type-2 Iuzzy set with 7 zSlices incorporating
the inIormation gathered Irom Ior example, Monday until
Sunday. In Iact, aIter the second day, a zSlices based general
type-2 set with 2 zSlices can be employed, the 3rd day one
with three zSlices and so Iorth.
Considering this case, the question arises on how to
update the resulting set on the Iollowing Monday - when a
new interval type-2 Iuzzy set becomes available. The
Iollowing choices are possible:
x Disregard the new inIormation. Disregarding the new
inIormation would make the system unable to adapt
to changing conditions or changing sensor/device
characteristics over time and thus is not desirable.
x Add an 8th zLevel and incorporate the new set. While
this is an option, increasing the number oI zLevels
results in a linear growth oI computational
perIormance requirements oI the Iuzzy system.
Additionally, the more interval type-2 Iuzzy sets are
incorporated, the larger generally the FOUs (at
individual zLevels), inhibiting crisp and accurate
response |16|.
x Regenerate the zSlices based general type-2 Iuzzy set
by replacing the interval type-2 Iuzzy set generated
on the previous Monday with the new interval type-2
set generated on this Monday. This approach is
successIul; nevertheless it prevents the system to
"home in" on an accurate model oI the device in
question: in this case the zSlices based set is always a
simple aggregation oI seven interval type-2 sets, it
does not Iocus on the continuing "agreement" oI the
individual sets over time.
x Incorporate the new interval type-2 Iuzzy set as a new
8th zSlice, and subsequently reduce the number oI
zLevels to seven by removing the zSlice at zLevel 1,
i.e.

1
. While this might seem arbitrary, it has a range
oI advantages which we are discussing below in the
rest oI this section.
By including the new interval type-2 Iuzzy set using the
process described in Section (III.A), we extract the new
inIormation Irom the interval type-2 set which directly
aIIects the individual zSlices making up the zSlices based
general type-2 Iuzzy set. An example oI this can be seen in
Section (IV).
Subsequently reducing the number oI zLevels to the
initially deIined (maximum) number oI zLevels by removing
zLevel

1
serves two purposes: Iirst, it prevents the Iuzzy
system Irom growing increasingly computationally complex.
As has been shown in |4|, |5|, the complexity oI evaluating a
zSlices based general type-2 Iuzzy system increases with the
number oI zLevels. Second,

1
, which is already associated
with the lowest zLevel (apart Irom zLevel 0 which is a



special case and can be disregarded as shown in |4|, |5|)
within the zSlices based general type2 Iuzzy set,
encompasses the largest FOU as it encompasses all the
FOUs oI each individual interval type-2 Iuzzy set. In other
words, its FOU includes the areas oI membership introduced
once or at least, the Iewest number oI times, typically by
"outlier" interval type-2 Iuzzy sets. In order to extract the
essential modelling inIormation in the zSlices based general
type-2 Iuzzy set, it is intuitive to rely on reoccurring
inIormation (i.e. multiple overlapping FOUs) and to discard
the inIormation included by outliers.
In other words, through removing

1
, we remove
inIormation which was less certain than that encoded in the
other zLevels. As such, consistent inIormation resulting
Irom "agreement" between source interval type-2 Iuzzy sets
is Iavoured over inIormation contributed by so called
outliers which are represented in

1
.
This process is repeated each day, maintaining a
consistent number oI zLevels. Thus, the process ensures a
continually adapting zSlices based general type-2 Iuzzy set
which dynamically adjust the FOUs oI its individual zSlices
to reIlect the best, most certain inIormation available. This in
turn allows Ior its application in continuously changing
conditions and environments while minimizing the tendency
Ior very large FOUs which interval type-2 Iuzzy systems are
oIten subject to and which can inhibit precise response |16|.
A worked example oI the process described here is provided
in Section IV.
IV. EXAMPLE
In order to better convey the concept and to illustrate the
process described in this paper, we are presenting a worked
example relying on triangular interval type-2 Iuzzy sets as
"inputs" to the process. The reason Ior choosing triangular
interval type-2 Iuzzy sets is two-Iold: Iirstly, as described in
Section (II.B), we are employing triangular interval type-2
Iuzzy sets as a basis Ior general type-2 Iuzzy sets in our
work in ambient intelligence and secondly, triangular Iuzzy
sets allow Ior a relatively easy graphical representation
which is essential Ior clarity, especially in terms oI general
type-2 Iuzzy sets as can be seen as part oI the Iigures in this
Section.
In the Iollowing example we are demonstrating the
process Ior the continuous integration oI a series oI interval
type-2 Iuzzy sets in order to generate a zSlices based general
type-2 Iuzzy set. In order to make the visualisation oI this
process possible, we have put a series oI restrictions on the
series oI Iuzzy sets presented here:
x The maximum number oI persistent zSlices was Iixed
at 2, growing to 3 only temporarily during the
integration oI an additional interval type-2 Iuzzy set.
In the ambient intelligence example mentioned
throughout the paper, the number oI persistent zSlices
is 7 (one per day), growing to 8 during the integration
oI another interval type-2 set (see Section III.B.), but
it is impossible to (visibly) represent such a large
number oI 3D zSlices as part oI a paper.
x The interval type-2 Iuzzy sets were chosen in order to
give rise to convex, normal zSlices based general
type-2 Iuzzy set in order to Iocus on the illustration oI
the process. Hence, no approximation and or
interpolation is required to assure convexity - this
greatly Iacilitates the visual identiIication oI the
individual sets throughout the process.
The three individual interval type-2 Iuzzy sets which are
employed are depicted in Fig. 5.


(a)

(b)

(c)
Fig. 5. (a) First Interval Type-2 Fuzzy Set. (b) Second Interval Type-2
Fuzzy Set. (c) Third Interval Type-2 Fuzzy Set.

In Fig. 5, the Iirst set describes the initial interval type-2
Iuzzy set available, i.e. in the context oI our ambient
intelligence example, it would be the set produced aIter the
Iirst day. The second set would consequently be produced on
the second day, while the third set would be produced on the
third day. As already mentioned, Ior visualisation reasons
we will limit our example to two permanent zSlices.
We are now proceeding by giving a step-by-step
description oI the continuous integration oI the individual
interval type-2 Iuzzy sets into a zSlices based general type-2
Iuzzy set. Each step is shown graphically in the
accompanying Iigures. Additionally, as it is inherently
diIIicult to show 3D objects such as general type-2
membership Iunctions in 2D, we have included interactive
3D representations oI Figs. 6-10 online here.




1. Initially we are considering the case where only the Iirst
interval type-2 Iuzzy set is available (Ior example aIter
the Iirst day). In terms oI zSlices based general type-2
Iuzzy sets it is equivalent to a zSlices based general
type-2 Iuzzy set with a single zLevel (i.e. interval type-
2 Iuzzy sets). A 3D representation oI the Iirst interval
type-2 set shown in Fig. 5 (a) is shown in Fig. 6.

(a) (b)
Fig. 6. Step 1, the Iirst interval type-2 set, shown in 3D. (a) From the
Iront/side. (b) From the rear/above.
2. As the second interval type-2 Iuzzy set (shown in Fig. 5
(b)) becomes available, the inIormation Irom the Iirst
and second interval type-2 Iuzzy sets are combined and
a two zSlices based general type-2 Iuzzy set is created
as shown in Section (III). The resulting zSlices based
general type-2 Iuzzy set is shown in Fig. 7. Considering
Fig. 7, it is visible that areas belonging to the FOUs oI
both the Iirst and the second interval type-2 Iuzzy set
have been assigned a secondary membership (the third
dimension shown on the zAxis) oI 1. Nevertheless,
areas which are covered by only one oI the initial two
membership Iunctions' FOUs are associated with a
secondary membership Iunction oI 1/2. The two
individual zLevels (at
1
= 1/2 and
2
= 1) employed
are a consequence oI employing two zLevels as shown
in Section (III.A) (Step 4 oI the process).

(a) (b)
Fig. 7. Step 2, zSlices based general type-2 set based on the Iirst and second
interval type-2 sets, shown in 3D. (a) From the Iront/side. (b) From the
rear/above.
3. The inclusion oI the third interval type-2 Iuzzy set (Fig.
5(c)) as shown in Fig. 8 illustrates the adaptation oI the
zSlices based general type-2 Iuzzy set as the number oI
interval type-2 Iuzzy sets exceeds the maximum number
oI zLevels speciIied (2 in this case, 7 in the example oI
weekdays given in Section (III)). As described in
Section (III.B), an additional, temporary zLevel is
created and the third interval type-2 Iuzzy set is
combined with the existing zSlices based general type-2
Iuzzy set. Fig. 8 shows how the zSlices based general
type-2 Iuzzy set is now based on three zLevels (at

1
= 1/3,
2
= 2/3 and
3
= 1). Comparing Fig. 7
and Fig. 8 illustrates how the inclusion oI the
inIormation encapsulated in the third interval type-2
Iuzzy set has aIIected the zSlices based general type-2
Iuzzy set.

(a) (b)
Fig. 8. Step 3, zSlices based general type-2 set based on the Iirst, second
and third interval type-2 sets, shown in 3D. (a) From the Iront/side. (b)
From the rear/above.
4. In order to maintain the speciIied maximum number oI
zLevels and more importantly, in order to allow Ior a
continuous adaptation oI the zSlices based general type-
2 Iuzzy set - and thus the model oI our speciIic device
or variable associated with this set, we now proceed to
remove the zSlice

1
. The Iirst part oI this process is
shown in Fig. 9, where

1
is identiIied (and shown as a
transparent wireIrame) beIore being removed. It is
important to note that

1
is not one oI the previous
interval type-2 Iuzzy sets, moreover it describes the area
which is associated with the lowest certainty (third
dimension / zLevel) oI the zSlices based general type-2
Iuzzy set. In other words, areas oI the FOU which have
been covered by the Iewest interval type-2 Iuzzy sets
(and thus can be considered as outliers) are removed
and the Iuzzy set is "pruned" to include only areas oI the
FOU which have been "seen" repeatedly and thus can be
associated with a higher level oI certainty. AIter

1
has
been removed, the number oI zLevels is returned to two
and the remaining zSlices are adjusted to reIlect this
(i.e.

3
(with a zLevel
3
= 1) becomes

2
(with a
zLevel
2
= 1) and

2
(with a zLevel
2
=
2
3
) becomes

1
(with a zLevel
1
=
1
2
)). The Iinal zSlices based
general type-2 Iuzzy set is shown in Fig. 10, while
comparing Fig. 9 and Fig. 10. visualises the change in
zLevels.


(a) (b)
Fig. 9. Step 4, removing zSlice 1(transparent) in order to reduce overall
number oI zLevels back to 2, shown in 3D. (a) From the Iront/side. (b)
From the rear/above.




(a) (b)
Fig. 10. Step 4, remaining 2 zLevels showing the resulting set and the
reduced FOU. (a) From the Iront/side. (b) From the rear/above.
With the process complete, the resulting zSlices based
general type-2 Iuzzy set provides a meaningful aggregation
oI the inIormation included in the individual interval type-2
Iuzzy sets. It can be said to be meaningful as the resulting set
is based on the concept oI overlapping (i.e. "agreeing") areas
oI the individual interval type-2 FOUs. Importantly, areas oI
the individual FOUs only overlap iI the source oI
inIormation which gave rise to the interval type-2 Iuzzy sets
has provided similar inIormation (a light sensor Ior example)
and thus does not represent a temporary Iluctuation or
outlier. II long term changes aIIect the characteristics oI a
device or variable, the described process will adapt the
general type-2 Iuzzy set.
V. CONCLUSIONS
In this paper we have proposed a process Ior generating
zSlices based general type-2 Iuzzy sets Irom a series oI
interval type-2 Iuzzy sets which was motivated by our work
in ambient intelligence. We have touched on the concept oI
"agreement" in terms oI overlapping FOUs oI multiple
interval type-2 Iuzzy sets which serves as a basis oI our
approach and have discussed how and why the aggregation
oI multiple interval type-2 sets can provide useIul
device/variable models Ior example in the context oI
ambient intelligence. In particular, we have discussed how
the proposed approach combats the tendency oI adapting
interval type-2 based systems which are subject to ever-
increasing FOUs and thus imprecise device/variable models
while providing the basis Ior a continually updated and thus
adaptive device/variable model.
While the concepts oI the process are straightIorward, we
have noted the complexity oI its implementation which is
largely related to the potential creation oI non-convex and
non-normal Iuzzy sets which in turn either requires the
establishment oI the appropriate theory in order to evaluate
such Iuzzy sets directly or the implementation oI
approximation mechanisms which convert the sets to their
convex (and potentially) normal counterparts.
We have provided a basic example oI the proposed
process using triangular Iuzzy sets, detailing the individual
steps oI the process and showing how the uncertainty model
oI the resulting zSlices based general type-2 Iuzzy set is
modiIied over time.
In the Iuture we are aiming to continue our investigation
into diIIerent implementations oI the proposed process as
part oI our work in ambient intelligence. While interpolation
and other approximation techniques are a possibility, we Ieel
that directly addressing the possibility oI computing with
non-convex general type-2 Iuzzy sets will be an important
aspect oI moving towards more accurate uncertainty
representations in general type-2 Iuzzy sets and thus better
perIormance oI general type-2 Iuzzy logic systems.
Finally, we would like to mention that the proposed
approach while based on interval type-2 Iuzzy sets, can also
be employed with type-1 Iuzzy sets as a source. We have
already investigated this possibility and we are aiming to
present the results as part oI Iuture publications.
Note: As many oI the Iigures in this paper are 3D and thus
are diIIicult to visualize on paper we have included
interactive online copies available here.
REFERENCES
|1| L.A. Zadeh, 'The Concept oI a Linguistic Variable and Its
Application to Approximate Reasoning-I, InIormation Sciences, vol.
8, no. 3, pp. 199-249, 1975.
|2| S. Coupland and R. John, 'Geometric type-1 and type-2 Iuzzy logic
systems, IEEE Transactions on Fuzzy Systems, vol. 15, pp. 3-15,
Feb. 2007.
|3| J. Mendel, Feilong Liu, 'On new quasi-type-2 Iuzzy logic systems,
Proceedings oI the IEEE International ConIerence oI Fuzzy Systems,
Hong Kong, China, June 2008, pp. 354-360.
|4| C. Wagner and H. Hagras, 'zSlices Towards bridging the gap
between Interval and General Type-2 Fuzzy Logic,' Proceedings oI
the IEEE International ConIerence oI Fuzzy Systems, Hong Kong,
China, June 2008, pp. 489-497.
|5| C. Wagner and H. Hagras, "Towards general Type-2 Fuzzy Logic
Systems based on zSlices," IEEE Transactions on Fuzzy Systems, to
be published.
|6| F. Liu, 'An eIIicient centroid type-reduction strategy Ior general type-
2 Iuzzy logic system, Information Sciences, Vol. 178, No. 9, pp.
22242236, May 2008.
|7| L. Lucas, T. Centeno and M. Delgado, 'General type-2 Iuzzy
inIerence systems: analysis, design and computational aspects,
Proceedings oI the IEEE International ConIerence oI Fuzzy Systems,
London, UK, July 2007, pp. 1107-1112.
|8| S. GreenIield and R. John, 'Optimised Generalised Type-2 Join and
Meet Operations, Proceedings oI the IEEE International ConIerence
oI Fuzzy Systems, London, UK, July 2007, pp. 141-146.
|9| S. Coupland, M. Gongora, R. John, and K. Wills, 'A Comparative
Study oI Fuzzy Logic Controllers Ior Autonomous Robots,
Proceedings oI IPMU , Paris, France, July 2006, pp. 13321339.
|10| C. Wagner and H. Hagras, 'zSlices based General Type-2 FLC Ior the
Control oI Autonomous Mobile Robots in Real World Environments,'
Proceedings oI the IEEE International ConIerence oI Fuzzy Systems,
Jeju Island, Korea, August 2009, pp. 718-725.
|11| C. Wagner and H. Hagras, 'Novel methods Ior the design oI general
type-2 Iuzzy sets based on device characteristics and linguistic labels
surveys, Proceedings oI the 2009 International Fuzzy Systems
Association World Congress and the 2009 European Society Ior
Fuzzy Logic and Technology ConIerence, pp. 537-543, Lisbon,
Portugal, July 2009.
|12| F. Doctor, H. Hagras, and V. Callaghan, 'A type-2 Iuzzy embedded
agent to realise ambient intelligence in ubiquitous computing
environments, Journal oI InIormation Sciences, vol. 171, no. 4, pp.
309-334, May 2005.
|13| J. Bezdek, "Pattern recognition with Iuzzy objective Iunction
algorithm," New York, Plenum, 1981.
|14| H. Hagras, "A Hierarchical type-2 Iuzzy logic control architecture Ior
autonomous mobile robots, IEEE Transactions On Fuzzy Systems,
vol. 12 No. 4, pp. 524-539, August 2004.
|15| H. Hagras, F. Doctor, V. Callaghan and A. Lopez, 'An incremental
adaptive liIe long learning approach Ior type-2 Iuzzy embedded agents
in ambient intelligent environments, IEEE Transactions on Fuzzy
Systems, Vol. 15, No. 1, pp. 4155, February 2007.
|16| C. Wagner, 'Towards better uncertainty handling based on zSlices
and general type-2 Iuzzy logic systems, PhD Thesis, University oI
Essex, June 2009.