A hybrid model based on constraint oselm, adaptive weighted src and knn for large scale indoor localization

Received December 6, 2018, accepted December 19, 2018, date of publication December 28, 2018,
date of current version January 23, 2019.
Digital Object Identifier 10.1109/ACCESS.2018.2890111
A Hybrid Model Based on Constraint OSELM,
Adaptive Weighted SRC and KNN for
Large-Scale Indoor Localization
HENGYI GAN 1, MOHD HARIS BIN MD KHIR1, (Member, IEEE),
GUNAWAN WITJAKSONO BIN DJASWADI1, AND NORDIN RAMLI2, (Senior Member, IEEE)
1Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, Seri Iskandar 32610, Malaysia
2Wireless Network and Protocol Research Laboratory, MIMOS Berhad, Kuala Lumpur 57000, Malaysia
Corresponding author: Hengyi Gan (heng_17005554@utp.edu.my)
This work was supported by the Universiti Teknologi PETRONAS under Grant 0153AA-H23.
ABSTRACT In this paper, a novel hybrid model based on the constraint online sequential extreme learning
machine (COSELM) classifier with adaptive weighted sparse representation classification (WSRC) and k
nearest neighbor (KNN) is proposed for the WiFi-based indoor positioning system. It is referred to as A Fast-
Accurate-Reliable Localization System (AFARLS). AFARLS exploits the speed advantage of COSELM
to reduce the computational cost, and the accuracy advantage of WSRC to enhance the classification
performance, by utilizing KNN as the adaptive sub-dictionary selection strategy. The understanding is
that the original extreme learning machine (ELM) is less robust against noise, while sparse representation
classification (SRC) and KNN suffer a high computational burden when using the over-complete dictionary.
AFARLS unifies their complementary strengths to resolve each other’s limitation. In large-scale multi-
building and multi-floor environments, AFARLS estimates a location that considers the building, floor, and
position (longitude and latitude) in a hierarchical and sequential approach according to a discriminative
criterion to the COSELM output. If the classifier result is unreliable, AFARLS uses KNN to achieve the best
relevant sub-dictionary. The sub-dictionary is fed to WSRC to re-estimate the building and the floor, while
the position is predicted by the ELM regressor. AFARLS has been verified on two publicly available datasets,
the EU Zenodo and the UJIIndoorLoc. The experimental results demonstrate that AFARLS outperforms the
state-of-the-art algorithms on the former dataset, and it provides near state-of-the-art performance on the
latter dataset. When the size of the dataset increases remarkably, AFARLS shows that it can maintain its
real-time high-accuracy performance.
INDEX TERMS Constraint online sequential extreme learning machine, k nearest neighbor, weighted sparse
representation classification, WiFi-based IPS.
I. INTRODUCTION
The implementation of the Internet of Things (IoT) is in a
wide diversity of fields ranging from smart infrastructure,
healthcare applications, industrial automation to real-time
monitoring and tracking, etc. The IoT-based object localiza-
tion and tracking are considered as one of the recent active
and immense developments of IoT applications [1]. Nonethe-
less, the localization and tracking system is not new. Since
the first satellite navigation system was studied by the U.S.
Navy using five satellites in 1960, massive developments
towards the system have continued even as it reached the
full operational status in 1995 [2]. Currently, the outdoor
geolocation of an object is obtained from the Global Posi-
tioning System (GPS) which uses four or more satellites.
Undeniably, the mature technology of the outdoor positioning
system that relies on the GPS has had a tremendous impact on
users’ everyday lives. Examples include navigation, tracking,
mapping and so forth. However, the GPS does not work well
in indoor environments because it requires line-of-sight (LoS)
measurement [3]. Hence, the precision of around 50 meters
achieved by the GPS for the non-line-of-sight (NLoS) of
reference objects in a complex indoor environment is very
limited to commercial applications [4]. To resolve these
limitations, various signals include WiFi, Bluetooth, RFID,
VOLUME 7, 2019
2169-3536
2018 IEEE. Translations and content mining are permitted for academic research only.
Personal use is also permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
6971

H. Gan et al.: Hybrid Model Based on COSELM, Adaptive Weighted SRC and KNN
ultrasound, light and magnetic field have been investigated
for IoT-based indoor positioning system (IPS).
The WiFi-based positioning technology is attracting
present-day scientific and enterprise interests owing to
its pervasive penetration of radio frequency signals and
extensive deployment of WiFi enabled devices [3], [5].
According to [6], there are two types of WiFi based position-
ing technologies. They are the time and space attributes of
received signal (TSARS)-based positioning technology and
the energy attribute of received-signal strength (RSS)-based
positioning technology. TSARS-based positioning technol-
ogy can be based on the techniques such as Time of
Arrival (ToA) [7], Angle of Arrival (AoA) [8] and Time
Difference of Arrival (TDoA) [9] of the WiFi signal to deter-
mine the users’ locations. AoA requires directional antennas
or antenna arrays to extract the information of the angle of
arrival signals. Meanwhile, ToA requires all the clocks of the
target and the anchor nodes to be in precise synchronization.
Unlike ToA, TDoA requires only the clock synchronization
for the anchor nodes [4]. Since RSS of the WiFi signal is sub-
jected to signal degradation as it traverses air over distances,
the RSS-based positioning technology can adopt a range-
based technique or a fingerprint-matching technique to locate
the users. Both techniques require no additional positioning
devices or clock synchronization. The range-based technique
that is based on the trilateration method depends highly on the
indoor radio propagation model [10], [11] and requires priori
precise location information of the APs (anchor nodes). Due
to it being very tough to well-establish a sophisticated indoor
radio propagation model in a dynamic indoor environment,
it suffers from low positioning accuracy. This difficulty arises
due to the radio signal propagation being mainly affected by
the interferences from the attenuation of the signal, multipath
fading and shadowing effects etc. [6], [12]. On the contrary,
the fingerprint-matching technique [13] requires no priori
location information of any of the APs. It comprises the
offline phase and the online phase [3], [6]. During the offline
phase, the intensity of the signal strengths of different APs
are collected with the MAC addresses at every location of
Reference Point (RP) to establish a radio map. In the online
phase, the real-time RSS obtained from the target node is
compared with the radio map via specific fingerprint-based
localization algorithms to estimate the most relevant position
of the target node. Although the fingerprint-matching tech-
nique achieves higher accuracy [14], it requires tremendous
setup and maintenance times, and different survey reduction
algorithms [3], [15].
The k nearest neighbor (KNN) algorithm is one of the sim-
plest and most effective among the fingerprint-based localiza-
tion algorithms in supervised machine learning. It works by
comparing the features of the testing data points with all the
labeled samples from a training dataset. According to some
prespecified distance metrics, the k nearest neighbors to the
points are extracted from the training dataset and contribute
equally for the final decision to label the points. In [16],
the best among 51 distance metrics have been investigated
with the KNN algorithm for the WiFi-based IPS. As a result,
the indoor positioning based on the Sorensen distance with
powed representation has been demonstrated to outperform
the traditional KNN methods, like the 1-NN based on raw
data and the most popular Euclidean distance. The main
advantage of this algorithm is that the training phase is very
minimal, apart from being simple and effective. Despite all
of these benefits, the time of computational complexity is
quite huge in the testing phase because of the necessity to
determine the distance of each query instance to all of the
training samples [17]. Further, it relies highly on the number
of the training samples. In other words, it performs better with
more reliable training samples, but it costs more time in the
operational computation and requires larger memory to store
the reference database. The other important factors that affect
the performance are the selection of the appropriate k value
and the decision rule in smoothing the k nearest neighbors.
Despite the WiFi-based IPS being able to establish high
positioning accuracy through fingerprint-based localization
algorithms like support vector machine (SVM) and KNN,
these traditional algorithms, which are batch learning meth-
ods, pose some disadvantages like huge labor-cost calibra-
tion, heavy time consumption in either the training or the
testing phases, and the recalibration required for differ-
ent environments [3]. Recently, extreme learning machine
(ELM) [18], [19] emerges as a very popular solution for large-
scale applications due to its extremely fast learning speed.
More advanced ELM algorithms are developed after that
and have been applied for better localization performance.
For instance, semi-supervised ELM (SELM) was proposed
in [20] to include graph Laplacian regularization so that it
did not depend too much on the labeled calibration data
for a location estimation. In [21], fusion semi-supervised
extreme learning machine (FSELM) was also proposed as
a better semi-supervised learning approach to reduce the
human calibration effort for indoor localization by consid-
ering the fusion information from WiFi and Bluetooth Low
Energy (BLE) signals. Moreover, online sequential extreme
learning machine (OSELM) [22] was developed and applied
for the indoor localization in [23] to address the problems
accordingly by using the traditional batch learning methods.
The fast learning speed of OSELM has proven that it can
help to reduce the intensive labor-cost and time-consuming
site survey during the offline phase. Besides that, its online
sequential learning ability can make the system to be more
invulnerable to environmental dynamics. For the sake of
the improved stability and generalization ability, the origi-
nal ELM was optimized with the L2 regularization param-
eter [24]. For the sake of clarity, we refer to the ELM that
uses L2 regularization as ELM-C in this paper. For better and
lifelong localization service, the L2 regularization parameter
was also introduced in OSELM, and it was referred to as
COSELM [25]. COSELM was developed to overcome the
fluctuation of the WiFi signal in the highly dynamic indoor
environments due to the changing status of the door [26],
the changing status of the relative humidity, and the people’s
6972 VOLUME 7, 2019

presence [27]. Therefore, it can maintain a viable system in a
longtime running performance. Although the COSELM clas-
sifier performs extremely fast, the original ELM itself handles
noise poorly. To make the classifier more robust against noise,
sparse representation classification (SRC) is of particular
interest to [28] by combining the ELM-C with the adaptive
SRC for image classification, referred to as EA-SRC. The
EA-SRC model is designed to inherit the respective excellent
characteristics of ELM which has low computational cost,
and SRC which has high prediction accuracy.
For long-lasting stable large-scale indoor localization in a
multi-floor single building environment, we have proposed
a novel indoor localization hybrid model, AFARLS, which
combines ELM, SRC and KNN. AFARLS estimates the
location that consists of floor and position (longitude and
latitude) in a hierarchical and sequential approach. Instead
of using the original ELM as a classifier, AFARLS employs
the COSELM classifier that utilizes the leave-one-out
(LOO) cross-validation scheme for the optimal regularization
parameter selection from [28]. Similar to the concept of
EA-SRC, AFARLS will only involve SRC if the classification
results in identifying the floor level are unreliable accord-
ing to a discriminative criterion to the COSELM output.
As mentioned previously, using an over-complete dictionary
for SRC poses a high computational complexity and lack
of adaptability. Thus, an adaptive sparse domain selection
strategy is very encouraged to resolve the negative effects of
uncorrelated classes of fingerprints. In AFARLS, we employ
a very different adaptive sparse domain selection strategy
for each query RSS fingerprint. We integrate AFARLS with
KNN based on the Sorensen distance metric with powed data
representation to achieve the best relevant sub-dictionary.
The sub-dictionary is subsequently fed to SRC to classify
the unreliable results again. Since the performance of the
KNN-based classification can be improved by introducing
a weighting scheme for the nearest neighbors, we develop
WSRC working with KNN to strengthen the classification
results as it is insufficient to distinguish the RSS fingerprint
through the residual alone. The noteworthy feature of WSRC
is based on the conceptual basics of weighted k nearest
neighbor (WKNN) that emphasizes close neighbors more
heavily. Rather than being based on the distances to the
query, WSRC is based on the residual to the query. Mean-
while, AFARLS utilizes the same sub-dictionary generated
from KNN to train ELM-C, and later to perform multi-
target regression to estimate the position. Dealing with the
multi-building and multi-floor indoor environments, the real-
world coordinates consist of the building, floor and posi-
tion (longitude and latitude) as a location. Accordingly, an
additional label of the building identification is considered in
COSELM classifier for multi-label classification. We verity
the proposed model in large-scale indoor environments with
two different databases, the EU Zenodo dataset [29] which is
a five-floor building with almost 22570m2 total surface area,
and the UJIndoorLoc dataset [30] which covers a surface
of almost 108703m2 including three buildings with either
four or five floors. Experimental results exhibit the real-
time high-accuracy localization performance of AFARLS
in a large-scale multi-building and multi-floor environment,
together with its long-term feasibility by leveraging online
incremental measurements to continuously update the model.
In short, the main contributions of this paper are as follows:
• We design a novel state-of-the-art localization algo-
rithm with an online sequential learning ability, combin-
ing COSELM that based on the LOO cross-validation
scheme and WSRC to complement each other in com-
putational complexity and classification accuracy.
• We propose COSELM as a novel clustering-based
approach in combination with KNN based on the
Sorensen distance metric with powed data representa-
tion as an adaptive sparse domain selection strategy to
achieve the best relevant sub-dictionary.
• We develop WSRC to emphasize close neighbors from
the sub-dictionary more heavily according to the resid-
ual to the query, rather than the distances to the query to
strengthen the classification performance.
The rest of this work is organized as follows. Section II
reviews the studies relevant to our work. Section III describes
and analyzes the system architecture. Section IV evaluates the
performance of the system model under large-scale indoor
localization environments. Finally, Section V concludes the
paper.
II. BACKGOUND INFORMATION
In this section, we review the studies pertinent to the frame-
work of our proposed system architecture, namely AFARLS.
The scope of the studies focuses on KNN, ELM and SRC
algorithms, together with their respective enhancements in
order to facilitate the understanding of our analytic model.
A. KNN
The KNN algorithm is widely utilized for classification,
though it can be used for regression. It extracts the k training
samples that are most identical or nearest to a test sample.
After that, the test sample is labeled based on the major-
ity label among the k nearest neighbors. Given a training
database, (xj, tj) ∈ Rn
× Rm
used in the supervised machine
learning, xj is the j-th n×1 training input vector (feature), tj is
the j-th m×1 training target vector (label) and j = 1, 2, . . . , N
samples. The output is to determine the k nearest neighbors
{r1, . . . , rk} in {xj}N
j=1 to an unlabeled test sample y ∈ Rn×1
,
and assign y based on the corresponding labels {tr1
, . . . , trk }.
In short, KNN classifies the pattern y to one of the possible
M classes of the problem at hand [31] as shown in (1).
y ← arg max
c∈{1,2,...,M}
[
k
X
l=1
Ic(rl)] (1)
where rl ∈ Rn×1
is the lth neighbor in x nearest to y according
to distancemetric (x, y), l = 1, 2, . . . ., k, c = 1, 2, . . . ., M,
and Ic(rl) is the indicator function of the lth nearest neighbor
VOLUME 7, 2019 6973

Algorithm 1 KNN-Based Classification
Input: A training data set {(xj, tj)}N
j=1
, a test sample y,
k parameter, and distancemetric (·).
Output: Class label of y
1: Compute the similarity rates between x and y according
to a distance metric
2: Select k nearest neighbors that have the highest similarity
rates
3: Label(y) = arg max
c∈{1,2,...,M}
[
Pk
l=1 Ic(rl)]
belonging to class c.
Ic (rl) =
(
1 if rl belongs to class c
0 otherwise
(2)
By introducing different weights to every nearest neighbor,
KNN is developed to WKNN. The weighting scale depends
on the distance between the test sample and the nearest neigh-
bor. The primary effect of the weighting scheme is that its
decision rule on the k nearest neighbors has less significance
on the majority vote than the closer neighbors. As a result,
the pattern y is assigned to one of the possible M classes in
which the rule is not merely based on the closest distance
to the query, but also the most frequent among its k nearest
neighbors, as shown in (3).
y ← arg max
c∈{1,2,...,M}
[
k
X
l=1
Ic(rl)Wrl ] (3)
Wrl =
1
distancemetric (y, rl)
(4)
Depending on the types of the distance metrics, the per-
formance of KNN is affected. Among the distance metrics
studied in [16], (5) is the most popular and it is also referred to
as the traditional KNN approach, while (6) has been recently
proven to be the most efficient for localization.
distanceeuclidean (x, y) = 2
v
u
u
t
n
X
i=1
|xi − yi|2 (5)
distancesorensen (x, y) =
Pn
i=1 |xi − yi|
Pn
i=1 (xi − yi)
(6)
distanceneyman (x, y) =
n
X
i=1
(xi − yi)2
xi
(7)
B. ELM
ELM is proposed based on a single-hidden layer feedfor-
ward neural network (SLFN) architecture [18], and the latter
extended to the generalized feedforward network. ELM itself
is a supervised batch learning method. Its notable merit lies in
it randomly selects the hidden node parameters (input weights
and bias), and then it determines only the output weights.
For N arbitrary distinct data samples, {(xj, tj)}N
j=1
is used in
the supervised batch learning, where xj = [xj1, xj2, . . . , xjn]
is a training input vector, n is the dimension of the feature,
tj = [tj1, tj2, . . . , tjm] is a training target vector, and m is
the dimension of the label. If a SLFN with additive L hidden
nodes can approach these N samples with no error, the output
function of the network is
fL xj

=
L
X
i=1
βiG wi, bi, xj

= sj
wi ∈ Rn
, bi ∈ R, xj ∈ Rn
, βi ∈ Rm
, sj ∈ Rm
i = 1, 2, . . . , L; j = 1, 2, . . . , N (8)
where wi = [wi1, wi2, . . . , win] is the input weight vector
connecting the ith hidden node to the input nodes and bi is
the bias of the ith hidden node, βi = [βi1, βi2, . . . , βim]T
is the output weight vector connecting the ith hidden node
to the output nodes, sj = [sj1, sj2, . . . , sjm] is the actual
network output vector with respect to xj, and G wi, bi, xj

is
the activation function. Since the hidden nodes are additive,
G wi, bi, xj

= g wi · xj + bi

(9)
Equation (8) can be summarized as
Hβ = T (10)
where,
H =



g (w1 · x1 + b1) · · · g (wL · x1 + bL)
.
.
.
...
.
.
.
g (w1 · xN + b1) · · · g (wL · xN + bL)



=



h (x1)
.
.
.
h (xN )



N×L
β =



βT
1
.
.
.
βT
L



L×m
, T =



tT
1
.
.
.
tT
N



N×m
(11)
Here H is called the hidden layer output vector. The notable
merit of the batch ELM algorithm is the random generation
of the hidden parameters of wi and bi without tuning during
training. Therefore, (10) becomes a linear system, and the β
is obtained by solving the following least squares problem to
minimize the error between tj and sj.
min
β
||Hβ − T|| (12)
Here, the β is estimated by
β̂ = H†
T (13)
where H† is the Moore-generalized inverse or the pseudo-
inverse of H, and it can be calculated as [18], i.e.
H†
=

HT
H
−1
HT
(14)
However, the pseudo-inverse suffers from numerical insta-
bility. Hence, ELM is enhanced to ELM-C by using the
regularized least squares from [32] to guarantee minimum
6974 VOLUME 7, 2019

Algorithm 2 ELM-C Based Classification
j=1
, a test sample y, acti-
vation functionG (·), and L hidden nodes.
1: Randomly generate parameters {wi, bi}
2: Calculate H
3: Compute λopt
4: Determine β̂
5: Compute the actual network output, s w.r.t y
c∈{1,2,...,M}
(s)
error. In ELM-C, the β is estimated by either (15) or (16).
When the training data sets are very large N ≥ L,
β̂ =

λoptI + HT
H
−1
HT
T (15)
else,
β̂ = HT

λoptI + HHT
−1
T (16)
where I is the identity matrix and λopt is the optimal regu-
larization parameter that can be determined from the range
of [−∞, ∞]. Since different regularization parameter λ gen-
erates different generalization performance, the LOO cross-
validation approach in [28] is adopted as an effective method
for the parameter optimization for the search of λopt, by con-
sidering a tradeoff between the training error term and λ.
Unless stated otherwise, λ ∈ [e−4, e4] is selected as the
candidate set of the regularization parameters in this paper to
determine λopt, and then the corresponding optimal β̂. This
optimization is beyond the paper’s scope.
The online learning version of ELM is OSELM. It consists
of two phases, an initialization phase and a sequential learn-
ing phase. The sequential learning phase updates the out-of-
date model with online incremental data that may come in
chunk-by-chunk or one-by-one. Given the initial large chunk
of data set, X0 = {(xj, tj)}
N0
j=1, where N0 is the number of the
initial training data in this chunk, the initial estimated β̂
0
is
β̂
0
= K−1
0 HT
0 T0 (17)
where,
K0 = HT
0 H0 (18)
When the new chunk of data set X1 = {(xj, tj)}
N0+N1
j=N0+1 arrives,
N1 is the number of new training data, the estimated β̂
1
is
β̂
1
= β̂
0
+ K−1
1 HT
1 (T1 − H1β̂
0
) (19)
where,
K1 = K0 + HT
1 H1, K0 = HT
0 H0 (20)
Consider p as the parameter that denotes the number of
updating times of the chunks of data presented for the online
sequential learning. In general, the solution of OSELM after
p + 1 times of incremental learning is
β̂
(p+1)
= β̂
(p)
+ K−1
p+1HT
p+1(Tp+1
− Hp+1β̂
(p)
) (21)
where,
Kp+1 = Kp + HT
p+1Hp+1, K0 = HT
0 H0 (22)
The stability and generalization performance of OSELM can
be improved in a similar way as performed for ELM by
utilizing the regularized least squares to replace the pseudo-
inverse. We refer it to as COSELM. In COSELM, given
the initial data X0 = {(xj, tj)}
N0
j=1, the initial estimated β̂
0
becomes
β̂
0
=

λoptI + HT
0 H0
−1
HT
0 T0 = K−1
0 HT
0 T0 (23)
where,
K0 = λoptI + HT
0 H0 (24)
and the estimated output weight of COSELM after p+1 times
incremental learning is
β̂
(p+1)
= β̂
(p)
+ K−1
p+1HT
p+1(Tp+1 − Hp+1β̂
(p)
) (25)
where,
Kp+1 = Kp + HT
p+1Hp+1, K0 = λoptI + HT
0 H0 (26)
Compared to OSELM, COSELM becomes OSELM when
λopt approximates 0, making the λoptI + HT
0 H0 ≈ HT
0 H0.
Hence, COSELM is a special case of OSELM, whereas
OSELM is a sepcial case of ELM. The extension of ELM
to COSELM trumps the traditional supervised batch learning
ELM, because it achieves better generalization performance
and is adaptive to changes.
C. SRC
Given a database with N arbitrary distinct data samples,
{Xc}M
c=1 for M classes, Xc ∈ Rn×kc that has n-dimensional
features and kc training input samples belonging to the class
c, the complete dictionary is formed as X = [X1, . . . , XM ].
In fact, the high dimensional and correlated training features
of X can be sparsely represented in sparse dictionary D that
contains a feature vector of much lower density. In general,
a sparse representation of X can be formalized as follows.
X = Dv + r (27)
where v represents the sparse coefficient vector and r denotes
the residual after sparse representation. To label a test sample
of Xt, the following two optimization problems can be used
to solve for the optimal v.
v̂ = arg min
v
kvk0 s.t. Dv = Xt (28)
v̂ = arg min
v
kvk1 s.t. kDv − Xtk2
2 ε (29)
where k·k0 is l0-norm, k·k1 is l1-norm and ε is a small error
tolerance in Xt. The latter is preferable for the minimization
VOLUME 7, 2019 6975

Algorithm 3 COSELM-Based Classification
j=1
, a test sample y, acti-
vation function G (·), L hidden nodes, initial chunk of data
size N0, subsequent chunks of online data size Np+1.
Initialization Phase
1: Randomly generate parameters {wi, bi}
2: Calculate H0
3: Compute λopt
4: Determine β̂
0
5: Set p = 0
Online Sequential Phase
6: If the (p + 1)th
chunk of data arrives, then
a: Calculate Hp+1
b: Update β̂
(p+1)
c: Set p = p + 1
d: Go to Step 6
7: else
a: Go to Step 9
8: end if
9: Compute the actual network output, s w.r.t y
c∈{1,2,...,M}
(s)
Algorithm 4 SRC-Based Classification
Input: A dictionary with M classes X = [X1, . . . , XM ], a test
sample Xt.
Output: Class label of Xt
1: Normalize the column vectors of X in unit of l2-norm
2: Solve the optimization problem (29)
3: Calculate the residuals, rc(Xt) =

Dδc(v̂) − Xt

2
2
, where
c = 1, 2, . . . , M
4: Label (Xt) = arg min
c∈{1,2,...,M}
rc (Xt)
problem in SRC due to the NP-hard problem of the for-
mer [28]. With the optimal v, the characteristic function δi(v̂)
in SRC generates a new vector that only possesses nonzero
coefficients in v̂ associated to the respective c-th class. If it
belongs to the class c, the residual of the class c, rc(Xt) =

Dδc(v̂) − Xt

2
2
is the minimum after it is approximately
represented by the sparse dictionary of that class; otherwise
the corresponding residual would be relatively huge.
Label (Xt) = arg min
c∈{1,2,...,M}
rc (Xt)
with rc(Xt) =

Dδc(v̂) − Xt

2
2
(30)
III. SYSTEM ARCHITECTURE
In this section, we present the architecture of the proposed
localization system model. AFARLS has three phases: offline
training phase, online sequential learning phase and online
localization phase or online testing phase. Three dimensional
real-world coordinates in which floor and position (longitude
and latitude) are considered as a location in a multi-floor
building. The fingerprint database D= {(xj, tj)}N
j=1
contains
N arbitrary data samples of WiFi fingerprints. The fingerprint
database is established through the popular collaborative or
crowdsourced method and collected using different mobile
devices. More precisely, the training input vector, xj =
[xj1, xj2, . . . , xjn] contains the fingerprints and the training
target vector, tj = [tj1, tj2, . . . , tjm] contains the physical
reference locations, where n is the number of APs present in
a building and m is the number of dimensions for a defined
location. Since the test bed is a single building, we set m = 3
for multi-label classification of the floor, the longitude and
the latitude. Before the offline training phase, the reference
locations are rounded to the nearest integers. The non-heard
AP fingerprints are represented with 0. The fingerprints in
which the RSS values are less than the RSS threshold, τ, are
also represented with 0. Next, the lowest RSS value, RSSmin,
is identified by considering all fingerprints and the APs of
the database. New data representation is performed according
to (31) based on [16]. As a result, the lowest value is 0 for
the non-heard or very weak-signal AP fingerprints, and the
higher values are for the stronger signals.
Positivei xj

= RSSi − (RSSmin − 1)
i = 1, 2, . . . , n; j = 1, 2, . . . , N (31)
During the offline training phase, the COSELM classifier
in AFARLS trains the fingerprint database at every reference
location. After the offline training phase, we can still update
the existing model during the online sequential learning phase
by leveraging the online incremental data. In the online
localization phase, given a test sample f ∈ Rn
, COSELM
classifies f to the location labels which are the estimated coor-
dinates that consist of the information of the floor level and
the position (longitude, latitude). These labels are classified
based on the three highest confident scores in a three-column
matrix of the actual network output vector, s ∈ Rm
. More
specifically, in a single building indoor environment, the first
column vector s1 and the second column vector s2 contain the
confident scores of the longitude and the latitude respectively,
while the last column vector s3 contains the confident scores
of the floor. If it is to predict the floor level associated to f ,
the floor is classified according to the highest confident score
in s3.
Since ELM itself is known to be less robust against noise in
classification tasks, the SRC-based classification is decided
based on a discriminative criterion of whether to be included
as an additional step to predict the floor level after COSELM.
The criterion measures the difference between the first and
the second highest confident scores in s3. We denote the
difference as δ3 = sF
3 − sS
3, where sF
3 and sS
3 are the first
and the second highest confident scores in s3 respectively. If
δ3 falls below an appropriate chosen positive threshold, σ,
the criterion is then satisfied and the misclassification rate of
COSELM under this condition tends to be higher. Therefore,
the floor level will not be estimated according to the highest
6976 VOLUME 7, 2019

confident score in s3. This is because of the fingerprint with
a small δ3 is contaminated with a larger noise signal that can
adversely affect the result more than the one with a big δ3. For
these noisy test fingerprints f , SRC is incorporated following
after COSELM.
Generating an adaptive sub-dictionary associated to f is
an important step for SRC to reduce the overall computation
burden. If the discriminative criterion is satisfied, the adaptive
sub-dictionary is used by SRC to predict the floor and is fed
to the ELM-C regressor to estimate the position. Otherwise,
it is used by ELM-C to predict the position. First, we alter
the usage of COSELM from classifying f to the location
labels to where the user belongs, to clustering a group of
relatively closely associated training fingerprints with the
corresponding reference locations. Instead of having one set
of the location labels as the final output, COSELM clusters
q pairs of position labels corresponding to the respective top
q highest confident scores in s1 and s2. In order to enhance
the probability of the classification rate, we increase the size
of the cluster group by establishing a new larger cluster
of the coordinates from the database whereby either their
longitude or latitude labels match to the q pairs of the position
labels. If the discriminative criterion is satisfied, the relevant
fingerprints, p associated to these coordinates are extracted
from the database. Otherwise, we further reduce the size
of the cluster by extracting only the relevant fingerprints
p, in which the corresponding coordinates have the floor
that matches the predicted floor determined from COSELM.
Although the size of the newly constructed subset has been
greatly reduced compared to the original database, the direct
application of the SRC-based classification to predict the
floor level and the direct application of ELM-C based regres-
sion to estimate the position are time-consuming. Hence, it is
appropriate to construct a fixed size of k, which is the smaller
subset of the extracted training samples. To address this issue,
we utilize the KNN-based classification as the adaptive sub-
dictionary selection strategy.
Despite employing KNN at the very beginning to deter-
mine the optimal k nearest neighbors from the database,
we propose it after COSELM. This can be explained in
terms of the reduced computational costs when KNN works
with clustering algorithms such as the K-means clustering
approach [33] and the support vector machine-based cluster-
ing approach [34]. Instead of using the initial large database,
KNN figures out the nearest neighbors from a cluster which
contains fewer samples which are more correlated to each
other. Since COSELM has demonstrated promising features
in computational speed and learning new data continuously,
we adopt COSELM as the clustering algorithm for KNN to
generate the adaptive sub-dictionary which contains the opti-
mal k nearest fingerprints, o to f . As pointed out previously,
the recent developed KNN based on the Sorensen distance
metric with powed data representation generates the best
result compared to the traditional KNN. Thus, we adopt KNN
that is based on this configuration. First, the highest RSS
value, RSSmax, is identified by considering the fingerprints, p
and f . Then, we convert the RSS values of p according to (32)
and return p̂.
Powedi (p) =
(p)α
(RSSmax)α = p̂, i = 1, 2, . . . , n (32)
where α is a mathematical constant, set to e. With p̂, KNN
based on the Sorensen distance metric is performed to gen-
erate the adaptive sub-dictionary that contains o in which the
RSS values are positive. For the sake of clarity, we refer to
these steps, starting from the output of COSELM to KNN,
as q-by-k clustering. After q-by-k clustering, we can per-
form the SRC-based classification to identify the floor by
using the adaptive sub-dictionary that contains the finger-
prints, o, provided that the discriminative criterion is satisfied.
Otherwise, the floor is directly determined from COSELM.
On the other hand, the position is estimated from ELM-
C based on the fingerprints in which the corresponding
coordinates belong to the predicted floor from the adaptive
sub-dictionary.
SRC sparsely represents parts of the nonzero elementary
features of the fingerprints to an adaptive sparse dictionary
that has much lower density for every floor, because of the
high-dimensional feature of the fingerprints is compressible.
For the floor level associated to f classified as the class
of c, f should be approximately represented by the sparse
dictionary of the c-th floor. In other words, for f that cannot be
approximately represented by a floor’s sparse dictionary, it is
considered not belonging to this floor. Since the performance
of KNN can be improved by introducing a weighting scheme
for the nearest neighbors, we develop WSRC working with
KNN based on the Sorensen distance metric with powed data
representation to strengthen the classification results as it is
insufficient to distinguish the RSS fingerprint through the
residual alone. The noteworthy feature of WSRC is that it
emphasizes close neighbors more heavily from the adaptive
sub-dictionary, similar to the basic idea of WKNN. Rather
than being based on the distances to the query, WSRC is
based on the residual to the query. At first, the corresponding
residual to each floor is converted to weightages. Based on
the k neighbors, the frequency to each floor’s vote is counted,
so that the majority vote of the neighbors is considered as the
metric for the classification accuracy. The primary effect of
the weighting scheme on the WSRC is that its decision rule on
the neighbors has less significance on the majority vote than
the neighbors that have the smallest residuals. As a result, f
is assigned to one of the possible floors in which the rule is
not merely based on the smallest residual to the query, but
also the most frequent among its k nearest neighbors. WSRC
classifies f to one of the possible M classes of the floors of
the problem at hand given as in (33).
Label (f ) = arg max
c∈{1,2,...,M}
[
k
X
l=1
Ic(ol)wc(f )] (33)
Ic (ol) =
(
1 if ol belongs to class c
0 otherwise
(34)
VOLUME 7, 2019 6977

Algorithm 5 WSRC-Based Classification
Input: A dictionary with k samples o = [o1, . . . , ok], a test
sample f .
Output: Class label of f
1: Normalize the column vectors of o in unit of l2-norm
2: Solve the optimization problem,
v̂ = arg min
v
kvk1 s.t. kDv − f k2
2 ε
3: Calculate the residuals, rc(f ) =

Dδc(v̂) − f

2
2
,
where c = 1, 2, . . . , M
4: Convert the residuals to weightages, wc(f )
5: Label (f ) = arg max
c∈{1,2,...,M}
[
Pk
l=1 Ic(ol)wc(f )]
wc(f ) =
1
rc(f )
(35)
where Ic(ol) is the indicator function of the lth fingerprint
in o corresponding to class c, wc is the c-th weightage
corresponding to the c-th residual, c = 1, 2, . . . , M, and
l = 1, 2, . . . , k. In short, we consider a hybrid classifier
model to predict the floor level in AFARLS which involves
COSELM, KNN and WSRC. On the other hand, the position
is determined from ELM-C after we have classified f to the
floor by using the fingerprints in which the corresponding
coordinates belong to the predicted floor from the adaptive
sub-dictionary o. Because of the sample size of o is relatively
small compared to the original database, the results from
the ELM-C and COSELM regressors are not much different.
Hence, we utilize ELM-C to train o, and compute the position
associated to f . LCOSELM and LELM−C are defined to denote
the number of hidden nodes for the COSELM classifier
and the ELM-C regressor respectively. Dealing with multi-
building and multi-floor indoor environments, we consider
a location has the real-world coordinates which consist of
the building, the floor and the position (longitude and lati-
tude). Accordingly, an additional label, which is the build-
ing identification, is required for multi-label classification in
COSELM (m = 4). The subsequent classification steps that
are customized to identify the building are very identical to
the hybrid classifier model that we have discussed before
to classify f to the floor. It is worth pointing out that it
will involve KNN and WSRC if δ4 from the output vector
s4 does not satisfy the criterion. To distinguish between the
thresholds, σ, belonging to the floor and the building, we use
σfloor and σbuilding. In summary, the detailed pseudo-code and
overview of AFARLS developed for a multi-building and
multi-floor indoor environment are presented in algorithm
6 and Figure 1.
IV. EXPERIMENT RESULTS AND DISCUSSION
This section presents the experimental results and discus-
sions. The results are obtained by running the experiments on
a PC, which has an Intel Core i7-4700MQ CPU at 2.40GHz
and 8GB RAM. Two large-scale publicly available Wi-Fi
crowdsourced fingerprint datasets are used to verify the actual
effect of the system model proposed in this paper in a multi-
floor single building indoor environment and a realistic multi-
floor multi-building indoor environment. The datasets are the
EU Zenodo database and the UJIIndoorLoc database.
The EU Zenodo database was created in 2017 at Tam-
pere University of Technology. It covers a five-floor building
that contains 822 rooms in total and has about 22570m2 of
footprint area of about 208m length and 108m width. The
database consists of a total of 4648 Wi-Fi fingerprints and the
corresponding reference locations. These locations contain
the floor levels, the longitude and the latitude coordinates (in
meters). From the database, it is split into the training set and
the testing set. The former and the latter contains 697 and
3951 WiFi fingerprints respectively. Each row of the WiFi
fingerprint is represented by a 992-column vector of MAC
addresses which contains the default value of +100dBm for
those non-heard APs. If the MAC address is received, then
it has a negative level of the RSS value (in dBm). All the
measurements were reported from 21 user devices which
could support both the 2.4GHz and 5GHz frequency range.
On the other hand, the UJIIndoorLoc database was created
in 2013 at Universitat Jaume I. It was reported from more
than 20 different users and 25 Android devices. It covers three
buildings with four or five floors. The total cover footprint
area is almost 108703m2. The database contains 21048 WiFi
fingerprints with the corresponding labels of the building,
the floor and the real-world longitude and latitude coordinates
(in meters) collected at pre-defined locations. Furthermore,
the database is split into the training set and the validation
set. The former contains 19937 training fingerprints and the
latter contains 1111 validation fingerprints. Since the publicly
available UJIIndoorLoc does not provide the testing sam-
ples which are made available only for the competitors at
EvAAL [30], we use the validation data as the testing data.
Each row of the WiFi fingerprints is represented by a 520-
column vector of MAC addresses which contains the default
value of +100dBm for those non-heard APs. If the MAC
address is received, then it has a negative level of the RSS
value (in dBm).
The core of the performance evaluation considers the train-
ing time, testing time, training accuracy and testing accu-
racy for the results. Throughout the experimentation process,
the time is reported on average by running it ten times,
whereas the positioning accuracy for the analysis of a pre-
dicted position (longitude and latitude) per test fingerprint
uses the Similarity function (in meters) in (38) based on
the Euclidean distance. To analyze N test fingerprints from
the validation database, three accuracy metrics (in meters)
are defined for evaluating the performance of the results.
First, the Error metric in (36) considers the two dimen-
sional (2D) mean positioning accuracy of the test fingerprints
regardless the correctness of building and floor identification.
On the other hand, the Error∗ metric considers the 2D mean
positioning accuracy of the test fingerprints whose building
and floor are correctly matched. Last, the Errorpn metric in
(37) considers the 2D mean positioning accuracy of the test
6978 VOLUME 7, 2019

FIGURE 1. Overview of AFARLS.
fingerprints by including the building and floor estimation
penalties based on the competition rule at EvAAL on IPIN
2015 [35].
Error =
1
N
N
X
i=1
Similarityi (x, y) (36)
Errorpn
=
1
N
N
X
i=1
Similarityi (x, y) + P1
× Bi + P2 × Fi (37)
Similarityi (x, y) =
q
(xi − x̃i)2
+ (yi − ỹi)2
(38)
where the Similarityi is based on the Euclidean distance
metric, (xi, yi) and (x̃i, ỹi) is the true and the estimated posi-
tions of the ith test fingerprint, P1 and P2 are the penalties
associated to wrong building and floor estimations, Bi is 1 for
wrong building estimations and 0 otherwise, Fi is the absolute
difference between the true floor and the estimated floor. P1
and P2 are set to 50 and 4 meters respectively. To evaluate the
classification accuracy in predicting the building and the floor
successfully, we define hit rate and success rate. The building
hit rate corresponds to the percentage of the test fingerprints
whose building is correctly predicted. The floor success rate
corresponds to the percentage of the test fingerprints whose
building and floor are correctly predicted. If we ignore the
correctness of the building identification, the floor success
rate becomes the floor hit rate that corresponds to the percent-
age of the test fingerprints whose floor is correctly predicted.
Hence, we consider the best localization metric is the one
in which its classification rate (highest hit rate or success
rate) is the highest, and then followed by the highest mean
positioning accuracy (lowest Error, Error∗ or Errorpn).
A. EXPERIMENT 1: SELECTION OF PARAMETERS FOR THE
COSELM CLASSIFIER IN AFARLS
Since COSELM is the crucial factor of our proposed model to
determine the overall performance, we investigate the critical
parameters that can affect its performance in this section with
the limited training fingerprints in the positive data represen-
tation based on the EU Zenodo database. We start by training
the classifier to identify the floor, while training the regressor
to estimate the position. Both the classifier and the regressor
are based on ELM. After that, the performance of the classi-
fication rate and the positioning accuracy using four different
activation functions are evaluated. These activation functions
VOLUME 7, 2019 6979

Algorithm 6 AFARLS
Input: Fingerprint database, D = {(xj, tj)|xj ∈ Rn
, tj ∈ Rm
}N
j=1, where m = 4 for multi-label classification, activation function
G (·), hidden node parameters (LCOSELM , LELM−C ), initial chunk of data size N0, subsequent chunks of online incremental
training data size Np+1, clustering parameters (q, k), threshold parameters (σbuilding, σfloor ), test fingerprint f .
Output: Predicted Coordinates, Z ∈ Rm
that considers building, floor and position (longitude, latitude) as a location
Pre-processing Phase:
1: Round off reference locations, t, to the nearest integers, return t
2: Represent non-heard and very weak-signal AP fingerprints in x with 0
3: Convert x to positive integers based on (31), return x
Offline Training Phase:
1: With {(xi, ti)|xi ∈ Rn
, ti ∈ Rm
}
N0
i=1, train the COSELM model based on algorithm 3
Online Sequential Learning Phase:
if the (p + 1)th
chunk of incremental data arrives, then
1: Update the COSELM model, with {(xi, ti)|xi ∈ Rn
, ti ∈ Rm
}
Np+Np+1
i=Np+1
end if
Online Localization Phase:
1: Generate s from the COSELM model, given f
2: Perform q-by-k clustering
2.1: Cluster q pairs of position labels corresponding to the respective top q confident scores in s1 and s2
2.2: Establish a cluster of coordinates whereby either their longitude or latitude labels match to these labels
2.3: if δ4 ≥ σbuilding then
a: Identify the building, Z4 by solving Label(f ) = arg max
c4∈{1,2,...,M4}
(s4)
b: Filter the cluster in which the building label of the coordinates is matched with Z4
2.4: end if
2.5: if δ3 ≥ σfloor then
a: Identify the floor, Z3 by solving Label(f ) = arg max
c3∈{1,2,...,M3}
(s3)
b: Filter the cluster in which the floor label of the coordinates is matched with Z3
2.6: end if
2.7: Extract p associated to these coordinates
2.8: Convert p to the powed data representation based on (32), return p̂
2.9: Determine o using KNN in algorithm 1 based on the Sorensen distance metric
3: if δ4 σbuilding then
a: Identify the building, Z4 with o based on WSRC in algorithm 5
b: Return o in which the corresponding building label is matched with Z4
4: end if
5: if δ3 σfloor then
a: Identify the floor, Z3 with o based on WSRC in algorithm 5
b: Return o in which the corresponding floor label is matched with Z3
6: end if
7: Use the ELM-C regressor in algorithm 2 with o to estimate the position (Z1, Z2)
8: Return Z = [Z1, Z2, Z3, Z4]
9: end
are the sigmoid function (sig), radial basis function (rbf), sine
function (sin) and hard-limit transfer function (hardlim).
Referring to Table 1, the top two best performances pre-
sented by the sig function and the hardlim function show that
they are more suitable to accept the raw data in positive data
representation. On the contrary, the performance of the RBF
is the worst. Since the performance and the training speed
using the sig function is better than the hardlim function,
it is selected as the activation function for the ELM approach
in IPS. Next, the original ELM is developed to ELM-C and
COSELM. Comparison of their performances is carried out
with the same activation function, and the initial number of
hidden nodes, L is selected to be 200. According to the results
in Table 2, the original ELM performs poorly compared
to the others. The performance improves substantially after
introducing the optimal parameter λopt into ELM, referred
to as ELM-C. Besides the activation function, the number
of the hidden nodes is another key parameter to determine
the performance of the positioning accuracy using the ELM
approach. L in the range of 200 to 5000 is studied for its effect
6980 VOLUME 7, 2019

TABLE 1. Selection of the type of activation function for ELM on the EU Zenodo database.
TABLE 2. Performance comparisons on the EU Zenodo database between ELM with and without λ and different number of L.
TABLE 3. Effects of the parameters, L and (N0, Np+1) on the ELM performance on the EU Zenodo database.
on the ELM-C performance. As observed in Table 2, the train-
ing time of ELM-C increases dramatically when the size is
1000 onwards, incremented with a step size of 2000. Despite
of the increase in time computation, both the training and
testing accuracies become relatively stable when the size
increases to 1000. In general, higher values of L generate
more stable and better performance, but it sacrifices its learn-
ing speed and it costs more time in the testing phase. Since
we will combine ELM with other algorithms for improve-
ment and consider the trade-off between the time and the
accuracy, we select the parameter L = 1000 as the opti-
mal condition for the experiments since it can generate the
results very close to the best within the acceptable computa-
tion time. Introducing online sequential learning method into
ELM-C, it becomes COSELM. The performance is evaluated
by splitting the 697 training samples into different pairs of
(N0, Np+1), together with the relationship between N0 and L.
As one can see in Table 3, the floor hit rate is the best when we
select N0 = 400, Np+1 = 50 and L = 1000, while the testing
Error∗ reduces to 11-12m when N0 L. The significant
increase in the floor hit rate and the positioning accuracy
can be explained when we increase L larger than N0, but it
costs more in the computational time. Meanwhile, it is noted
that the sequential learning method can speed up the learning
process only when N0 ≥ L. Therefore, the appropriate selec-
tion of L and N0 is crucial for the accuracy performance and
the time complexity tradeoff. With a larger training database,
the online sequential learning method becomes more essen-
tial to help in reducing the training time. We can realize its
noteworthy speed advantage in Experiment 3 when we utilize
a bigger training database and set N0 ≥ L.
B. EXPERIMENT 2: PERFORMANCE EVALUATION BASED
ON THE EU ZENODO DATABASE
The design parameters in AFARLS are configured as follows
on the EU Zenodo database, unless stated otherwise. Sig is
selected as the activation function, σfloor is 0.2, the number of
hidden nodes of COSELM and ELM-C (LCOSELM , LELM−C )
are 1000 and 100 respectively, the clustering parameter pair
(q, k) are 8 and 4 respectively, and τ is set to −103dBm. The
training dataset of 697 samples is split into the initial data
chunk size N0 of 400, followed by the subsequent multiple
data chunks with the size Np+1 of 50.
The performance comparisons between AFARLS and
the benchmark positioning algorithms on the EU Zenodo
database is reported in Table 4. In this table, we see AFARLS
VOLUME 7, 2019 6981

FIGURE 2. Performance comparisons between AFARLS and the KNN approaches on the EU Zenodo database during online testing phase. (a) Floor Hit
Rate vs k Nearest Neighbors. (b) Erro r∗ vs k Nearest Neighbors.
outperforms other algorithms. In the meantime, we list the
best results from different approaches of KNN with their
respective k compared with AFARLS in Table 5. As shown
in the table, AFARLS presents the best performance by uti-
lizing less computational time compared to the best result
among all KNN approaches. The outstanding performance of
AFALS in a single-building indoor environment is because
of it has KNN and WSRC, which work together to further
enhance the COSELM results. We also test the performance
of AFARLS with and without the help of KNN and WSRC.
Shown in the same table, AFARLS uses only COSELM and
skips the algorithms of WSRC and KNN when q = 0 and
k = 0. As a result, despite of it reduces the testing time
significantly, its performance drops heavily. It is also noted
that the training time of AFARLS remains approximately the
same with and without the sequential learning method; this
is because the training database is not very huge, whereby
N0 is smaller than L. Different adaptive sub-dictionary sizes
which contain different numbers of nearest neighbors, k and
q = 8 are studied to evaluate their performances. Their
performances are compared with the KNN approaches by
varying k from 1 to 15. Figure 2(a) relates the effect of k
to floor hit rate in AFARLS and compares AFARLS with
the performances using KNN approaches. On the other hand,
Figure 2(b) describes the effect of the number of k to testing
Error∗. As depicted in both figures, AFARLS trumps the
KNN algorithms regardless of the factor of k, and it generates
the results very fast for the real-time applications.
C. EXPERIMENT 3: PERFORMANCE EVALUATION BASED
ON THE UJIIndoorLoc DATABASE
The design parameters in AFARLS are configured as follows
on the UJIIndoorLoc database, unless stated otherwise. Sig
is selected as the activation function, σbuilding is 0.2, σfloor is
TABLE 4. Performance comparisons between AFARLS and the benchmark
positioning approaches on the EU ZENODO database.
0.8, the number of hidden nodes of COSELM and ELM-C
(LCOSELM , LELM−C ) are 1000 and 100 respectively, the clus-
tering parameter pair (q, k) are 8 and 6 respectively, and τ
is set to −104dBm. The training dataset of 19937 samples
is split into the initial data chunk size N0 of 2000, followed
by the subsequent multiple data chunks with the size Np+1
of 500.
Table 6 reports the results from the recent existing bench-
mark positioning algorithms on the UJIIndoorLoc database.
As discussed before, the testing data is only provided to
the competitors at EvAAL. Therefore, a direct performance
comparison between AFARLS and the first four positioning
algorithms from the EvAAL/IPIN 2015 competition is not
6982 VOLUME 7, 2019

TABLE 5. Performance comparisons between AFARLS with the KNN approaches on the EU ZENODO database.
TABLE 6. Performance comparisons between AFARLS and the benchmark
positioning approaches on the UJIIndoorLoc database.
possible. On the other hand, the fifth positioning algorithm
validates the result by splitting the training data into new
training and validation sets with the ratio of 70:30. The
last positioning algorithm adopts the performance evaluation
method similar to ours by treating the UJIIndoorLoc vali-
dation data as the testing samples. Even though the direct
comparison is not allowed, our results are very comparable to
the results presented in Table 6. Meanwhile, we list the best
results from different approaches of KNN with their respec-
tive k compared with AFARLS in Table 7. Summarized in the
table, AFARLS makes the best classification performances
in building hit rate and floor success rate with the help of
WSRC and KNN, whereas its testing Error∗ is the second-
lowest. Furthermore, it is noteworthy that the testing time
of AFARLS is reported to be the lowest. This is because
AFARLS benefits from the speed advantage of ELM. With
the help of the online sequential learning algorithm, the train-
ing time can also be reduced to almost half. The same as with
Experiment 2, we also test the performance of AFARLS with
and without the help of KNN and WSRC. When q = 0 and
k = 0, AFARLS utilizes only COSELM and skips the algo-
rithms of WSRC and KNN. As expected in Table 7, the per-
formance of AFARLS drops heavily using only COSELM,
but the testing time has been saved significantly. To validate
the effect of different number of nearest neighbors k to the
performance in AFARLS and to compare it with the per-
formances using the KNN approaches, we vary k from 1 to
15 and set q = 8. Figure 3(a) and Figure 3(b) describe the
effect of k on building hit rate and floor success rate respec-
tively, while Figure 3(c) depicts the effect of k on the testing
Error∗. From these figures, we see that AFARLS outperforms
many KNN algorithms in classifying the building and the
floor, while the position estimated from AFARLS is around
0.2-0.3m different compared to the best performance among
all KNN algorithms. Nevertheless, the unique advantage of
AFARLS over KNN is that it can reduce significantly the
cost of online testing time. Once trained, it can estimate the
locations faster than KNN given the test fingerprint database
because it benefits from the speed advantage of the ELM
algorithm. In addition, the online sequential learning method
introduced to ELM helps AFARLS in speeding up the offline
training process especially when we deploy it in a larger
environment with a bigger training database. As a result,
AFARLS saves remarkable training and testing time.
VOLUME 7, 2019 6983

FIGURE 3. Performance comparisons between AFARLS and KNN on the
UJIIndoorLoc database during online testing phase. (a) Building Hit Rate
vs k Nearest Neighbors. (b) Floor Success Rate vs k Nearest Neighbors. (c)
Error∗ vs k Nearest Neighbors.
Apart from being fast in the training and testing phases,
AFARLS can update itself by using the online incremen-
tal data so that it can adapt itself to a new and different
environment without retraining the whole system. To show
this feature with the UJIIndoorLoc dataset, we sort its training
and test samples according to the building label. Hence, we
have three sets of training datasets and three sets of test
datasets. The first set of training and test datasets corre-
sponds to the samples that belong to the first building, while
the second and third datasets correspond to the samples that
belong to the second and third buildings respectively. During
the offline training phase, we establish an initial AFARLS
model with the first training dataset. After that, we evaluate
the system performance by using the testing samples that
belong to the first test dataset. Next, we update the existing
model during online sequential learning phase by using the
subsequent chunk of dataset in which all the training samples
belong to the second training dataset. Then, we evaluate the
system performance with the first and the second test datasets.
Last, we update the model again by using the third dataset
and evaluate the system performance by using all the test
datasets.
As illustrated in Table 8, AFARLS performs very well as
expected in a single building environment. This is because the
test bed is small. When the environment is enlarged to include
additional new buildings (the second and third buildings),
the performance declines slightly. It should also be noted
that these results are very similar to the results we have
obtained in the previous section when we train the system
with all the training data fully prepared in advance during
the offline training phase. In fact, preparing all the data in
advance is labor-cost and time-consuming. It is very hard to
collect all the required training data ahead of time. There-
fore, AFALRS makes use of the online sequential learning
method to suit the way the new training data arrives without
sacrificing the accuracy and it adapts itself to a new envi-
ronment without retraining a new model with all the training
data.
To show the impact of the online incremental data on
the system performance, we randomly select 10000 samples
from the training dataset as the online incremental data to
reflect the environmental dynamics. In other words, we set
up an initial model in AFARLS with the initial training
data of 9937 samples during offline training phase. During
the online sequential learning phase, we update the existing
model with five chunks of datasets sequentially in which
each dataset has 2000 incremental data. Then, we evaluate
the performance of AFARLS after every update to the model
with a step size of 2000 samples. Figure 4(a) depicts the
influence of the online incremental data to the system per-
formance in the online localization phase, after every update
to the model. As shown, the initial localization performance
is the worst. After gradually refining the model with the
online incremental data, the performance improves because
the newly updated model can reflect the current indoor envi-
ronment better. This tells us that AFARLS can improve
and update itself along the timeline by utilizing the incre-
mental learning method for a lifelong and high-performance
running.
6984 VOLUME 7, 2019

FIGURE 4. (a) Performance evaluation of AFARLS with respect to multiple chunks of online incremental data on the UJIIndoorLoc database.
(b) Misclassification comparisons of AFARLS with respect to different σ on the EU Zenodo and the UJIIndoorLoc databases.
TABLE 7. Performance comparisons between AFARLS and the KNN approaches on the UJIIndoorLoc database.
D. EXPERIMENT 4: PERFORMANCE EVALUATION OF
AFARLS WITH DIFFERENT THREDHOLDS AND
PARAMETERS
In this experiment, the threshold σ is tested on the grid
[0:0.1:1]. It is to demonstrate the variations of building hit
rate and floor hit rate after partitioning the test fingerprint
database into subsets based on σ. For the test fingerprints
classified as the noisy fingerprints, AFARLS incorporates
the classification capability of WSRC to identify the floor.
Following the analysis in Figure 4(b), the highest floor hit
rate in the experiment based on the EU ZENODO database is
achieved when σfloor = 0.2, while the best floor hit rate in the
experiment based on the UJIIndoorLoc database is achieved
when σfloor = 0.8. In addition, it can be realized that the result
based on the former experiment does not have much gain
corresponding to the change of σ, compared with the latter.
It can be explained in such a way that the fingerprint database
of the former experiment contains little noise when it is used
to classify the floor, whereas the latter experiment carried
out in the larger (multi-building) environment contains the
noisier fingerprint database when it is used to classify the
floor but the noise is distinguishable when the database is
used to classify the building. It can be seen from Figure 4(b)
that the building hit rate is stable for σbuilding ∈ [0, 1.0].
VOLUME 7, 2019 6985

FIGURE 5. Misclassification and testing accuracy comparisons of AFARLS with respect to various combinations of q and k during online
testing phase. Tested on the EU Zenodo database: (a) Floor Success Rate vs k Nearest Neighbors. (b) Error∗ vs k Nearest Neighbors. Tested
on the UJIIndoorLoc database: (c) Floor Success Rate vs k Nearest Neighbors. (d) Error∗ vs k Nearest Neighbors.
TABLE 8. Performance evaluation of AFARLS with respect to three
training UJIIndoorLoc datasets sorted according to the building label.
Therefore, we consider a small threshold σfloor = 0.2 as
the optimized condition in the former experiment, and utilize
σbuilding = 0.2 for the building identification, but adopt
σfloor = 0.8 which has a larger noise tolerance for the floor
identification in the latter experiment. Next, we realize that
the choice of the clustering parameter pair (q, k) determines
the adaptive sub-dictionary size. Larger values of q and k pro-
duce bigger size of the sub-dictionaries, followed by a higher
cost of computational complexity. Figure 5(a) and Figure 5(b)
plot the testing results in the experiments based on the EU
ZENODO database and Figure 5(c) and Figure 5(d) plot the
testing results in the experiments based on the UJIIndoorLoc
database with respect to various combinations of q and k,
respectively. The floor success rate is reported as 94.59%
on average, and the testing Error∗ is reported as 7.74m on
average in the regions of q ∈ [8, 20] and k ∈ [4, 15] on the
EU Zenodo database, depicted in Figure 5(a) and Figure 5(b).
On the other hand, the floor success rate is reported as 95.22%
on average, and the testing Error∗ is reported as 6.38m on
average in the regions of q ∈ [6, 20] and k ∈ [3, 15]
depicted in Figure 5(c) and Figure 5(d). It is worth pointing
out that larger values of q and k do not guarantee better
performance, but they will lead to a stable performance. For
instance, we can see that the regions where q ∈ [2, 4] and
k ∈ [1, 2] in both experiments produce very low accuracy
results. The worst cases that happen in both experiments have
6986 VOLUME 7, 2019

been discussed previously in Table 5 and 7 when q = 0 and
k = 0, the results drop dramatically though the computation
cost are extremely small. Hence, the proper selection of these
parameters is vital for generating the appropriate size of the
sub-dictionary that contains sufficient informative data to
work out accurate results within the acceptable computational
time. In view of all experiments above, AFARLS exploits
the speed advantage of ELM to reduce the computational
cost, and the accuracy advantage of SRC to enhance the
classification performance, by utilizing KNN as the adaptive
sub-dictionary selection strategy. The baseline results that are
based on the traditional KNN are 91.72% for the floor hit rate
and 8.88m for the testing Error∗ on the EU Zenodo dataset.
On the other hand, the baseline results that are based on the
traditional KNN are 99.55% for the building hit rate, 89.92%
for the floor success rate and 7.90m for the testing Error∗
on the UJIIndoorLoc dataset. Compared to our proposed
model, the results from AFARLS are 94.76% for the floor hit
rate and 7.58m for the testing Error∗ on the former dataset,
while they are 100% for the building hit rate, 95.41% for
the floor success rate and 6.40m for the testing Error∗ on
the latter dataset. Thus, AFARLS has enhanced the floor hit
rate by 3.31% and the testing Error∗ by 14.64% based on
the former dataset, while the building hit rate is enhanced
by 0.45%, the floor success rate by 6.11% and the testing
Error∗ by 18.99% based on the latter dataset. As one can
see, the computational cost of the traditional KNN approach
based on the latter dataset (i.e., 84.01s) is very large compared
to the former dataset (i.e., 17.52s) during the online testing
phase. The intensive computational cost in KNN is on account
of the huge training sample size of the latter dataset. Unlike
KNN, AFARLS demonstrates its outstanding computational
efficiency via the adaptive sub-dictionary selection strategy.
When the size of the dataset increases remarkably, AFARLS
shows its testing time will not increase significantly and it can
estimate the results faster than the KNN algorithms. More
specifically, it saves up to almost 70% of the testing time
compared to the traditional KNN approach based on the latter
dataset. Not only is the testing speed improved, AFARLS also
adopts the online sequential learning method to speed up the
training speed. Apart from exploiting the online sequential
learning ability for the speed advantage, AFARLS utilizes this
method to update the existing model in a timely manner to
environmental dynamics with online incremental data. Most
importantly, AFARLS learns fast with a varying chunk size
of the new incremental data. Without retraining a new model
with all the training data to update the system, AFALRS
cuts down the time consumptions and manpower costs for
the site survey during the offline training phase. Considering
these unique advantages, AFARLS outrivals the traditional
fingerprinting-based positioning algorithms based on KNN.
V. CONCLUSION
In large-scale highly dynamic indoor environments, the out-
standing WiFi-based IPS should offer advantages of not
merely the high positioning accuracy, but also the lifelong
performance running, together with the fast and feasible site
survey. In this paper, we have proposed AFARLS to offer
these advantages in the large-scale indoor environments and
have validated the performance through extensive experi-
mentation. The results show that AFARLS can offer real-
time performance with high accuracy, and leverage online
incremental data with a varying size to update the out-of-date
model without retraining a new model. This novel adaptive
Wi-Fi indoor localization model inherits the advantages from
the original ELM, SRC and KNN algorithms in order to
tackle their respective drawbacks which render their practical
applications in IPS. Future improvement can be focused on
filtering algorithms and more advanced ELM algorithms to
enhance the performance and reduce the memory size of the
reference database by storing only the reliable training finger-
prints collected through the popular collaborative or crowd-
sourced method.
ACKNOWLEDGMENT
The author would like to thank great appreciation towards
MIMOS for the hardware assistance.
REFERENCES
[1] S. Ramnath, A. Javali, B. Narang, P. Mishra, and S. K. Routray, ‘‘IoT based
localization and tracking,’’ in Proc. Int. Conf. IoT Appl. (ICIOT), 2017,
pp. 1–4.
[2] SGM Caspari, ‘‘Global positioning systems in combat,’’ U.S.
Army Sergeants Major Acad., Fort Bliss, TX, USA, Tech. Rep.
Accessed: Apr. 22, 2018. [Online]. Available: https://apps.dtic.mil/dtic/
tr/fulltext/u2/a596592.pdf
[3] S. He and S.-H. G. Chan, ‘‘Wi-Fi fingerprint-based indoor positioning:
Recent advances and comparisons,’’ IEEE Commun. Surveys Tuts., vol. 18,
no. 1, pp. 466–490, 1st Quart., 2015.
[4] R. F. Brena, J. P. García-Vázquez, C. E. Galván-Tejada, D. Muñoz-
Rodriguez, C. Vargas-Rosales, and J. Fangmeyer, Jr., ‘‘Evolution of indoor
positioning technologies: A survey,’’ J. Sensors, vol. 2017, pp. 1–21,
Mar.2017, Art. no. 2630413.
[5] R. Henniges, ‘‘Current approaches of Wi-Fi positioning,’’ in Proc.
IEEE Conf. Publications, 2012, pp. 1–8. Accessed: Apr. 22,
2018. [Online]. Available: https://www.snet.tu-berlin.de/fileadmin/
fg220/courses/WS1112/snet-project/wifi-positioning_henniges.pdf
[6] S. Xia, Y. Liu, G. Yuan, M. Zhu, and Z. Wang, ‘‘Indoor fingerprint
positioning based on Wi-Fi: An overview,’’ ISPRS Int. J. Geo-Inf., vol. 6,
no. 5, p. 135, 2017.
[7] R. Kaune, ‘‘Accuracy studies for TDOA and TOA localization,’’ in Proc.
15th Int. Conf. Inf. Fusion (FUSION), 2012, pp. 408–415.
[8] J. Xu, M. Ma, and C. L. Law, ‘‘Cooperative angle-of-arrival position
localization,’’ Measurement, vol. 59, pp. 302–313, Jan. 2015.
[9] Y. Yan, H. Wang, X. Shen, K. He, and X. Zhong, ‘‘TDOA-based source
collaborative localization via semidefinite relaxation in sensor networks,’’
Int. J. Distrib. Sensor Netw., vol. 11, no. 9, p. 248970, 2015.
[10] G. Retscher, ‘‘Fusion of location fingerprinting and trilateration based on
the example of differential Wi-Fi positioning,’’ ISPRS Ann. Photogramm.,
Remote Sens. Spatial Inf. Sci., vol. 4, p. 377, Sep. 2017.
[11] J. Zhang, G. Han, N. Sun, and L. Shu, ‘‘Path-loss-based fingerprint
localization approach for location-based services in indoor environments,’’
IEEE Access, vol. 5, pp. 13756–13769, 2017.
[12] R. Javaid, R. Qureshi, and R. N. Enam, ‘‘RSSI based node localiza-
tion using trilateration in wireless sensor network,’’ Bahria Univ. J. Inf.
Commun. Technol., vol. 8, p. 58, Dec. 2015.
[13] G. Félix, M. Siller, and E. N. Álvarez, ‘‘A fingerprinting indoor localization
algorithm based deep learning,’’ in Proc. 8th Int. Conf. Ubiquitous Future
Netw. (ICUFN), 2016, pp. 1006–1011.
[14] H. Chen, Y. Zhang, W. Li, X. Tao, and P. Zhang, ‘‘ConFi: Convolutional
neural networks based indoor Wi-Fi localization using channel state infor-
mation,’’ IEEE Access, vol. 5, pp. 18066–18074, 2017.
VOLUME 7, 2019 6987

[15] F. Ebner, T. Fetzer, F. Deinzer, and M. Grzegorzek, ‘‘On Wi-Fi model opti-
mizations for smartphone-based indoor localization,’’ ISPRS Int. J. Geo-
Inf., vol. 6, no. 8, p. 233, 2017.
[16] J. Torres-Sospedra, R. Montoliu, S. Trilles, Ó. Belmonte, and J. Huerta,
‘‘Comprehensive analysis of distance and similarity measures for Wi-Fi
fingerprinting indoor positioning systems,’’ Expert Syst. Appl., vol. 42,
pp. 9263–9278, Dec. 2015.
[17] M. Zhao and J. Chen, ‘‘Improvement and comparison of weighted K
nearest neighbors classifiers for model selection,’’ J. Softw. Eng., vol. 10,
no. 1, pp. 109–118, 2016.
[18] G.-B. Huang, Q.-Y. Zhu, and C.-K. Siew, ‘‘Extreme learning machine:
Theory and applications,’’ Neurocomputing, vol. 70, nos. 1–3,
pp. 489–501, 2006.
[19] G. Huang, G.-B. Huang, S. Song, and K. You, ‘‘Trends in extreme
learning machines: A review,’’ Neural Netw., vol. 61, pp. 32–48,
Jan. 2015.
[20] J. Liu, Y. Chen, M. Liu, and Z. Zhao, ‘‘SELM: Semi-supervised ELM with
application in sparse calibrated location estimation,’’ Neurocomputing,
vol. 74, no. 16, pp. 2566–2572, 2011.
[21] X. Jiang, Y. Chen, J. Liu, Y. Gu, and L. Hu, ‘‘FSELM: Fusion semi-
supervised extreme learning machine for indoor localization with Wi-Fi
and Bluetooth fingerprints,’’ Soft Comput., vol. 22, no. 11, pp. 3621–3635,
2018.
[22] N.-Y. Liang, G.-B. Huang, P. Saratchandran, and N. Sundararajan, ‘‘A fast
and accurate online sequential learning algorithm for feedforward net-
works,’’ IEEE Trans. Neural Netw., vol. 17, no. 6, pp. 1411–1423,
Nov. 2006.
[23] H. Zou, X. Lu, H. Jiang, and L. Xie, ‘‘A fast and precise indoor localiza-
tion algorithm based on an Online sequential extreme learning machine,’’
Sensors, vol. 15, no. 1, pp. 1804–1824, Jan. 2015.
[24] G.-B. Huang, H. Zhou, X. Ding, and R. Zhang, ‘‘Extreme learning machine
for regression and multiclass classification,’’ IEEE Trans. Syst., Man,
Cybern. B, Cybern., vol. 42, no. 2, pp. 513–529, Apr. 2012.
[25] Y. Gu, J. Liu, Y. Chen, and X. Jiang, ‘‘Constraint online sequential extreme
learning machine for lifelong indoor localization system,’’ in Proc. Int.
Joint Conf. Neural Netw. (IJCNN), 2014, pp. 732–738.
[26] Y.-C. Chen, J.-R. Chiang, H.-H. Chu, P. Huang, and A. W. Tsui, ‘‘Sensor-
assisted Wi-Fi indoor location system for adapting to environmental
dynamics,’’ in Proc. 8th ACM Int. Symp. Modeling, Anal. Simulation
Wireless Mobile Syst., 2005, pp. 118–125.
[27] K. Kaemarungsi and P. Krishnamurthy, ‘‘Properties of indoor received
signal strength for WLAN location fingerprinting,’’ in Proc. 1st Annu. Int.
Conf. Mobile Ubiquitous Syst., Netw. Services (MOBIQUITOUS), 2004,
pp. 14–23.
[28] J. Cao, K. Zhang, M. Luo, C. Yin, and X. Lai, ‘‘Extreme learning machine
and adaptive sparse representation for image classification,’’ Neural Netw.,
vol. 81, pp. 91–102, Sep. 2016.
[29] E. S. Lohan, J. Torres-Sospedra, H. Leppäkoski, P. Richter, Z. Peng, and
J. Huerta, ‘‘Wi-Fi crowdsourced fingerprinting dataset for indoor position-
ing,’’ Data, vol. 2, no. 4, p. 32, 2017.
[30] J. Torres-Sospedra et al., ‘‘UJIIndoorLoc: A new multi-building and multi-
floor database for WLAN fingerprint-based indoor localization problems,’’
in Proc. Int. Conf. Indoor Positioning Indoor Navigat. (IPIN), 2014,
pp. 261–270.
[31] M. Bicego and M. Loog, ‘‘Weighted K-nearest neighbor revisited,’’
in Proc. 23rd Int. Conf. Pattern Recognit. (ICPR), 2016,
pp. 1642–1647.
[32] M. van Heeswijk and Y. Miche, ‘‘Binary/ternary extreme learning
machines,’’ Neurocomputing, vol. 149, pp. 187–197, Feb. 2015.
[33] B. Altintas and T. Serif, ‘‘Improving RSS-based indoor positioning algo-
rithm via k-means clustering,’’ in Proc. 11th Eur. Wireless Conf.-Sustain.
Wireless Technol. (Eur. Wireless), Apr. 2011, pp. 1–5.
[34] C.-W. Lee, T.-N. Lin, S.-H. Fang, and Y.-C. Chou, ‘‘A novel clustering-
based approach of indoor location fingerprinting,’’ in Proc. IEEE 24th
Int. Symp. Pers. Indoor Mobile Radio Commun. (PIMRC), Sep. 2013,
pp. 3191–3196.
[35] A. Moreira, M. J. Nicolau, F. Meneses, and A. Costa, ‘‘Wi-Fi fingerprinting
in the real world-RTLS@ UM at the EvAAL competition,’’ in Proc. Int.
Conf. Indoor Positioning Indoor Navigat. (IPIN), 2015, pp. 1–10.
[36] A. Razavi, M. Valkama, and E.-S. Lohan. (2015). ‘‘K-means fingerprint
clustering for low-complexity floor estimation in indoor mobile localiza-
tion.’’ [Online]. Available: https://arxiv.org/abs/1509.01600
[37] A. Cramariuc, H. Huttunen, and E. S. Lohan, ‘‘Clustering benefits in
mobile-centric WiFi positioning in multi-floor buildings,’’ in Proc. Int.
Conf. Localization GNSS (ICL-GNSS), 2016, pp. 1–6.
[38] S. Shrestha, J. Talvitie, and E. S. Lohan, ‘‘Deconvolution-based indoor
localization with WLAN signals and unknown access point locations,’’ in
Proc. Int. Conf. Localization GNSS (ICL-GNSS), 2013, pp. 1–6.
[39] J. Torres-Sospedra et al., ‘‘A realistic evaluation of indoor positioning
systems based on Wi-Fi fingerprinting: The 2015 EvAAL–ETRI competi-
tion,’’ J. Ambient Intell. Smart Environ., vol. 9, no. 2, pp. 263–279, 2017.
[40] K. S. Kim, S. Lee, and K. Huang, ‘‘A scalable deep neural network
architecture for multi-building and multi-floor indoor localization based
on Wi-Fi fingerprinting,’’ Big Data Anal., vol. 3, p. 4, Dec. 2018.
[41] M. Nowicki and J. Wietrzykowski, ‘‘Low-effort place recognition with
WiFi fingerprints using deep learning,’’ in Proc. Int. Conf. Autom., 2017,
pp. 575–584.
HENGYI GAN received the B.S. degree in elec-
trical and electronic engineering from Universiti
Teknologi PETRONAS (UTP), in 2017, where he
is currently pursuing the M.S. degree. He is cur-
rently a Research Associate with the Department
of Electrical and Electronic Engineering, UTP.
His research interest includes indoor localization,
wireless sensor networks, machine learning, and
artificial intelligence.
MOHD HARIS BIN MD KHIR (M’08) was born
in Kedah, Malaysia, in 1968. He received the
B.Eng. degree in electrical and electronic engi-
neering from Universiti Teknologi MARA, Selan-
gor, Malaysia, in 1999, the M.Sc. degree in com-
puter and systems engineering from Rensselaer
Polytechnic Institute, NY, USA, in 2001, and the
Ph.D. degree in systems engineering from Oakland
University, MI, USA, in 2010. He joined Univer-
siti Teknologi PETRONAS (UTP), in 1999, where
he is currently an Associate Professor with the Electrical and Electronic
Engineering Department. He held several positions at UTP such as the
Deputy Head of Department and the Director of mission oriented research
on nanotechnology. Most devices were fabricated using CMOS and MUMPS
technologies. He has published three book chapters and more than 37 jour-
nals and 70 conference paper in the area of sensor, actuator, energy harvester,
and sensor’s application in the IoT. His research interest include micro/nano–
electro mechanical systems sensors and actuator development.
GUNAWAN WITJAKSONO BIN DJASWADI
received the B.S. (magna cum laude) and M.S.
degrees in electrical engineering from Michigan
Technological University, Houghton, MI, USA,
in 1992 and 1994, respectively, and the Ph.D.
degree in electrical and computer engineer-
ing from the University of Wisconsin–Madison
in 2002. From 1994 to 1996, he was with the
National Aeronautics and Space Agency, Indone-
sia. In 2002, he joined Denselight Semiconduc-
tors Pte Ltd., Singapore, where he developed high-speed, long wavelength,
and distributed feedback lasers. He was with Finisar Malaysia to develop
uncooled and high-speed optical transceiver. He was with the Department of
Electrical Engineering, University of Indonesia, from 2005 to 2007, before
joining MIMOS when he held various key positions such as a Principal
Researcher and the Director of Research and Sensor System Architect,
until 2016. He is currently an Associate Professor with the Electrical and
Electronics Department, Universiti Teknologi PETRONAS, where he is also
an Indonesia-Chapter Professional Engineer.
6988 VOLUME 7, 2019

NORDIN RAMLI (M’04–SM’13) received the
B.Eng. degree in electrical engineering from Keio
University, Japan, in 1999, and the M.Eng. and
Ph.D. degrees in electronic engineering from
the University of Electro-Communications, Japan,
in 2005 and 2008, respectively. He was with
Telekom Malaysia, Berhad, as a Network Engi-
neer, from 1999 to 2008, and a Lecturer with Mul-
timedia University, Malaysia, from 2008 to 2009.
He is currently a Senior Staff Researcher with
Wireless Network and Protocol Research, MIMOS Berhad, Malaysia. He is
also a Solution Architect for the Internet of Things (IoT) and big data related
project. He has authored or co-authored over 80 journals and conference
papers, and has filed over 30 patents related to wireless communications
which are pending with World Intellectual Property Organization and the
Intellectual Property Corporation of Malaysia. His current research interests
include cognitive radio, TV white space, space-time processing, equaliza-
tion, adaptive array system and wireless mesh networking, the IoT, and
big data. He has been appointed as a member of the Young Scientist Net-
work of Malaysia Academy of Science, since 2014. He received the Top
Research Scientist Malaysia, in 2018. He has been the Chair of the White
Space Working Group, a technical standardization working group, Malaysia
Technical Standard Forum, Berhad, since 2013, to study and promote the
technology of white space communication in Malaysia. He is currently an
Associate Editor of IEICE Communication Express. He is also a Registered
Professional Engineer with the Board of Engineers, Malaysia.
VOLUME 7, 2019 6989

A hybrid model based on constraint oselm, adaptive weighted src and knn for large scale indoor localization

More Related Content

What's hot (20)

Similar to A hybrid model based on constraint oselm, adaptive weighted src and knn for large scale indoor localization (20)

More from Journal Papers (20)

Recently uploaded (20)

A hybrid model based on constraint oselm, adaptive weighted src and knn for large scale indoor localization