extreme learning machine for user location prediction in mobile environment
TRANSCRIPT
Extreme learning machinefor user location prediction
in mobile environmentTeddy Mantoro
Department of Computer Science, KICT,International Islamic University Malaysia,
Kuala Lumpur, Malaysia
Akeem OlowolayemoDepartment of Information Systems, KICT,
International Islamic University Malaysia, Kuala Lumpur, Malaysia
Sunday O. OlatunjiDepartment of Computer Science, FICT, Universiti Teknologi Malaysia (UTM),
Johor Bahru, Malaysia, and
Media A. Ayu and Abu Osman Md. TapDepartment of Information Systems, KICT,International Islamic University Malaysia,
Kuala Lumpur, Malaysia
Abstract
Purpose – Prediction accuracies are usually affected by the techniques and devices used as wellas the algorithms applied. This work aims to attempt to further devise a better positioning accuracybased on location fingerprinting taking advantage of two important mobile fingerprints, namely signalstrength (SS) and signal quality (SQ) and subsequently building a model based on extreme learningmachine (ELM), a new learning algorithm for single-hidden-layer neural networks.
Design/methodology/approach – Prediction approach to location determination based onhistorical data has attracted a lot of attention in recent studies, the reason being that it offers theconvenience of using previously accumulated location data to subsequently determine locations usingpredictive algorithms. There have been various approaches to location positioning to further improvemobile user location determination accuracy. This work examines the location determinationtechniques by attempting to determine the location of mobile users by taking advantage of SS andSQ history data and modeling the locations using the ELM algorithm. The empirical results show thatthe proposed model based on the ELM algorithm noticeably outperforms k-Nearest Neighborapproaches.
Findings – WiFi’s SS contributes more in accuracy to the prediction of user location than WiFi’s SQ.Moreover, the new framework based on ELM has been compared with the k-Nearest Neighbor and theresults have shown that the proposed model based on the extreme learning algorithm outperforms thek-Nearest Neighbor approach.
Originality/value – A new computational intelligence modeling scheme, based on the ELM has beeninvestigated, developed and implemented, as an efficient and more accurate predictive solution fordetermining position of mobile users based on location fingerprint data (SS and SQ).
Keywords Location awareness, Artificial neural network, Extreme learning machines,Mobile technology, Learning
Paper type Research paper
The current issue and full text archive of this journal is available at
www.emeraldinsight.com/1742-7371.htm
IJPCC7,2
162
Received 14 December 2010Revised 6 January 2011Accepted 16 March 2011
International Journal of PervasiveComputing and CommunicationsVol. 7 No. 2, 2011pp. 162-180q Emerald Group Publishing Limited1742-7371DOI 10.1108/17427371111146446
1. IntroductionThe prime goal of pervasive computing is the idea of information technology (IT)services every time and everywhere (Weiser, 1991; Stanton, 2001). This makesIT-enabled services available and responsive to user’s needs in mobile environmenttaking advantages of their location and context information. This implies that there isno gain saying about the importance of location and context aware computing owningto its pivotal role in context-based services management.
There have been wide areas of applications of context computing and locationawareness with a rapidly growing demand and research for its applications in variousnew areas. Some of the application areas include resources optimization (e.g. power,bandwidth, etc.), fleet tracking and shipping applications, underground as well asunderwater navigations, security – tracking criminals by law enforcement officials,automotive applications, advertisement and marketing, aged and disabled supportsystems, robot and machine control and a host of all others (Deitel, 2001; Mautz, 2008).
Obviously, there is an urgent need for better approaches that will improve theapplications of mobile positioning techniques, with adequate and up-to-date devices aswell as appropriateness of algorithms in ways that will carefully analyze the existingtechniques and methods to be able to develop sound strategies to further enhancecontext aware applications.
Through context computing, IT infrastructures and services can now be madeavailable and better utilized by adapting to users’ needs and thus facilitating enhancedresources management.
Prediction accuracies are usually affected by the techniques and devices used aswell as the algorithms applied (Mantoro and Olowolayemo, 2009). The appropriatemanagement of these components of context determination is more often than not oneof the major problem areas in location positioning systems.
The methodology adopted in this paper is to review existing positioning systemsand approaches, location referencing types, technologies, as well as differentpositioning algorithms. Subsequently, our approach at further improving our initialresults by modeling with extreme learning machine (ELM) algorithm is reported.
The work discusses previous approaches to location recognition based onfingerprinting of signal attributes peculiar to each position in the experimentallocation using wireless local area network (WLAN) signals. This work is an attempt atimproving previous positioning accuracy based on location fingerprinting takenadvantage of another important mobile fingerprints, namely signal quality (SQ) inaddition to signal strength (SS) that has been used extensively in previous works, suchas in (Kaemarungsi, 2006; Rohrig and Kunemund, 2007; Kumar et al., 2006) andsubsequently building a discovery model based on ELM, a new learning algorithm forsingle-hidden-layer neural networks (SLFNs) described in Huang et al. (2004). In thesubsequent section each of the components of positioning systems are furtherelaborated.
The remaining part of this paper is organized as follows. The next section gives areview of related work on key concepts in location and context awareness, locationfingerprinting and neural networks (NN) with special focus on ELM. Section 3discusses the system’s design techniques and the last section; conclusion, evaluates theresults of the study as well as the challenges encountered and also proposes our futuredirections.
ELM for userlocation
prediction
163
2. Related workLocation determination is the most important component of context awareness (Mantoroand Olowolayemo, 2009). There have been tremendous attempts previously to betterapproximate location of mobile users. These are due largely to the ever-growing needfor better positioning for improved location-based services’ management. Differentapproaches have been used and proposed in literatures. One such approach that hasgained more recognition recently is location fingerprinting. Location fingerprinting is toposition users based on differential signal attributes at different locations rather thancomputing the distance between the signal transmitting points, usually the access point,and mobile device terminal peculiar to other network-based approaches (Reyero andDelisle, 2008). The received SS is then compared to location history data previously storedin a radio map. Efforts at using location fingerprinting to determine mobile user locationshave been explored in several previous works such as in Tsai et al. (2009), Mantoro andOlowolayemo (2009), Mantoro et al. (2009), Kaemarungsi and Krishnamurthy (2004)and Kaemarungsi (2006). Specifically, in Tsai et al. (2009), a comparison of NN modelsusing SS was explored. The NN models were based on the simulated-annealingback-propagation and feedforward neural network (FFNN) with back-propagation(BPN), showing better performance results of the latter over that of the former models.
2.1. Positioning levels of orientationsPositioning systems accuracies can be tailored to three levels of orientation. Theseorientations are point-, line- and area-orientation (Brimicombe and Li, 2009). Locationpositioning systems’ accuracies is an issue of great concern especially, when thepositioning has to do with point-orientation, such as those services meant for disabledand aged citizens who require assistance. In such cases, the accuracy must be tailoredto the “exact” position of the person-concerned. However, in this and other orientationssuch as line or area-orientations, there is still need for improvement to give betterconfidence and better management of resources. A good example is the case of wrongpositioning of fleet management (Brimicombe and Li, 2009).
2.2. Positioning approachesPositioning approaches can be classified into three; network-based, device-based andhybrid methods, comprising both first two approaches (Brimicombe and Li, 2009).
Network-based positioning often regarded as integrated approach since the network isused for communication and data transmission to other networks as well as determinationof user location. In network-based positioning, user locations are determined byestimating the signal travelling to and from a set of transmitter base stations, such as,WLAN access points (APs) or global system for mobile communications (GSM) basestations. The position is, thereby measured through computation of the length anddirection of the radio path of a mobile device from these base stations (Brimicombe andLi, 2009). Network-based positioning methods do not require extra software or hardwareto be installed in the mobile device since the network is used for the computation. Power isalso relatively more conserved unlike in device-based methods that require high-power forcomputation. Network-based approaches are known to be more accurate since, they haveaccess to fine grained data which include transmission power, antenna location, and tilt;which are not available in device-based systems (Reyero and Delisle, 2008). There aremany network-based positioning techniques that are used in indoor users’ location
IJPCC7,2
164
techniques for measuring signals at a particular location to an Access Point. Theseinclude cell of origin (the earliest method of locating a wireless user by determining thetransmitting range where the call is made), angle of arrival (AOA) – locating a user’sposition by the overlapping of two cell where the cell phone receive signal, enhancedobserved time difference – determining user’s position by measuring the time travelbetween the phone and multiple towers, time of arrival (TOA), time difference of arrival(TDOA) – locating a user by determining the time, it takes the signal from the cell-phone toreach the transmitting tower. The other issue in device-based positioning is that there is anadded overhead of computation power as a result of processing of location information onthe device while at the same time being used for network (Brimicombe and Li, 2009;Mantoro and Olowolayemo, 2009).
However, for outdoor users’ location determination global positioning system (GPS)is used. It is the de facto standard for outdoor positioning. GPS uses satellites to track auser’s latitude, longitude and altitude, it can be used virtually everywhere in the world,including on airplanes and ship (Brimicombe and Li, 2009). The main drawback of GPSis that it is not effective for indoor location determination. Recently, assisted-GPS, thatis, a combination of GPS with one or more of network-based positioning approaches,such as Cell-ID, TDOA, AOA, has been used as hybrid approaches to work around thisshortcoming of GPS (Brimicombe and Li, 2009).
2.3. Location fingerprintingRecently, location fingerprinting, that is, recalling patterns (such as multipath) whichmobile phone signals are known to exhibit at different locations in each cell, has beenproposed and adjudged as better suited for indoor positioning. However, in locationfingerprinting positioning, the appropriate selection of location determination algorithmsand models for context enabled services has been very crucial for accuracy andenhancement of positioning systems and applications, which is one of the keys to futuremobile infrastructures and services management and security. Various indoor approachessuch as those used in location fingerprinting give some advantages over device-basedmethod that uses GPS. This is due to two important reasons. One, GPS performsinadequately in indoor environment due to obstruction to signal and non-line of sight,which lead to attenuation and poor SQ while location fingerprinting on the other hand isnot affected by non-line of sight. Out of these methods, location fingerprinting has beenconsidered a more appropriate choice. This is because other techniques are affected bynon-line of sight and multi-path in indoors space (Mantoro and Olowolayemo, 2009). Thisleads to errors and inaccuracies in the training datasets captured.
2.4. SensorsPositioning sensors are classified based on the level of precision to which they candetermine a mobile location. Broadly, there are Precise, Proximate and Predictedapproaches to level of precision positioning.
Precise sensors. Location estimation techniques that are based on sensors that candetermine a mobile user location to ,1 m accuracy. Examples in this category includeradio frequency identification, etc.
Proximate sensors. Location estimation techniques based on sensors that candetermine a mobile user location up to a meter accuracy. Examples of sensors that fallin this category include WiFi, GSM, WiMedia, ZigBee, active/passive badge, voicerecognition, face recognition, smart floor (Mantoro and Johnson, 2003).
ELM for userlocation
prediction
165
Predicted approaches. In this approach, historical data of mobile user location iscollected, stored in a repository which can then be used to estimate mobile user locationby pattern matching, using different types of models that are broadly classified asdeterministic, based on Euclidean or Manhattan norm, such as the k-Nearest Neighborand probabilistic, such as maximum-a-posterior or maximum-likelihood frameworks(Honkavirta et al., 2009).
2.5. PrinciplesThe principles for location determination can be broadly classified under proximity,trilateration or triangulation, multilateration and centroids. The fundamental principleas decribed in Reyero and Delisle (2008) is discussed below.
Proximity, the most primitive of the principles, rely on the transmitting tower’srange when a call is made from a cell phone. It is the least accurate since the user can beanywhere within a specific area or cell (Deitel, 2001; Figure 1).
Trilateration and triangulation use the overlap of towers cells to compute thelocation of a user. If a couple of towers receive signals from a phone their cells can beused to calculate the location of the phone. Trilateration, the basic principle underlyingcommon positioning technologies (Huang et al., 2010), uses distance or a measure ofdistance, inferred from other attributes such as SS, TOA, etc. to determine userlocation. The determination of mobile user location is computed from the intersectionof cells or signals from at least three sources. Solving the system of equations belowdescribed in Reyero and Delisle (2008) gives the location coordinates (Figure 2).
Triangulation uses signal AOA from at least two reference points, then determinesthe location as intersection of lines (Reyero and Delisle, 2008). The solution to thelocation coordinate is given by the system of equations below (Figure 3).
Multilateration, the technique finds the location of mobile users by preciselycomputing the TDOA of a signal emitted from the device to at least three receivers. It isoften referred to as hyperbolic positioning.
In centroid, a user location is calculated as an aggregated weighted mean of thedetected reference points’ location, where the weight could be calculated from the receivedSS, the reference point coverage, or the reliability of the reference points (Figure 4).
The linear centroid can be generally expressed as:
x ¼i
Xwixi; y ¼
i
Xwiyi
Figure 1.Cell of origin,the most primitivelocation principle
Cell-phone user canbe anywhere withinthe tower's range.
Source: Deitel (2001)
IJPCC7,2
166
The accuracy of these techniques improves as the number of reference points increases,the distance of the points decreases and with increase in the number of other factorsincluded (Reyero and Delisle, 2008).
2.6. AlgorithmsFurthermore, there are different algorithms that could be used to determine the userlocation. The major component of context awareness system that determines the accuracyof the predictions is the appropriateness and adequacy of the algorithm used. There are anumber of major algorithms that have been adopted or presented in context awarenesstexts. These include Markov models, Hidden Markov models (Wallbaum and Spaniol,2006; Nuno-Barrau and Paez-Borrallo, 2005), Bayesian networks (Fox et al., 2003),
Figure 2.Trilateration
y
x
(k,1)
(i,j)
(0,0)
Note: {r12 = x2 + y2, r2
2 = (x – i)2 + (y – j)2,[(r13)]r2 = [(x – k)]r2 + [((y – l))]r2}Source: Reyero and Delisle (2008)
r2
r1r3
Figure 3.Triangulation
(k,1)
(i,j)
(0,0) x
y
α
β
Note: y = (x–i)sin(a) + j, y = (x–k)sin(b) + lSource: Reyero and Delisle (2008)
ELM for userlocation
prediction
167
NN (Tsai et al., 2009) or time-series. In several previous works, location hasbeen estimated using a number of algorithms and techniques. The algorithms forlocation estimations can be classified as deterministic and probabilistic (Honkavirta et al.,2009). Under deterministic approach, which is based on the traditional Euclideandistance approximation method, approaches with different norms were considered.The probabilistic approach is generally based on the conditional probability distributionfunction using Bayes’ rule.
2.7. Proposed methodologyThe technique proposed in this work is to estimate users’ location based on locationfingerprinting specifically collecting mobile devices location data, namely, the SS and SQ, asfirst proposed in Mantoro et al. (2009), based on proximity to a number of APs.Subsequently, a model based on ELM modeling is used to estimate the location of the mobileuser, a process regarded as place detection in some literatures (Reyero and Delisle, 2008).
2.8. Artificial neural networkThe prediction technique used is discussed in this section. This involve the developmentand application of ELM, a new artificial neural network (ANN) framework which isderived from SLFNs (Huang et al., 2000), to the determination of location of mobile usersbased on the SS and SQ peculiar to the particular location. Our discussion opens with ANNto gain full understanding of the underlying principles of SLFNs and ELM.
ANN is a mathematical model that tries to stimulate the functions of the brain(Tan et al., 2006). Similar to the human brain, the neural network is composed ofan interconnected assembly of nodes and directed links. It is known that the humanbrain learns by changing the strength of the synaptic connection between neuronsupon repeated stimulation by the same impulse (Bernacki and Wlodarczyk, 2005).A generic structure of the traditional ANN is shown in Figure 5.
The aim is to compute the value of the computed output y close enough to the targetoutput f(x).
The ANN is defined, thus.Given a finite input nodes x ¼ xi; . . . ; xk with corresponding weight vector,
w ¼ wi; . . . ;wk the network’s response, y is given by:
Figure 4.Centroid
x(0,0)
(x1,y1)
(x3,y3)
w3 w4
w2w1
(x4,y4)
(x2,y2)
yIJPCC7,2
168
y ¼ gX
kwk · xk
� �[ ð0; 1Þ
adjusting the weights sufficiently to estimate y, after a finite number of steps, suchthat:
y < f ðxÞ
by minimizing the difference between the network’s response, y and the desiredoutput, f(x). The termination condition is achieved after a specified number ofiterations, usually referred to as epoch is reached. The performance of ANN isoften specified in form of root mean square error (RMSE). The network is known tousually converge, however, optimal convergence is not guaranteed.
In order to accelerate convergence in multilayered feedforward neural network(MFN), BPN is often used. The goal is to learn the weights, thereby minimizing themean squared error. This implies that we first compute:
okj ¼ gjX
iwji · okj
� �
for each hidden and output unit.Subsequently, for each node uj, the accumulated error dkj:
dkj ¼ ðtkj 2 okjÞg0j
Xiwji · oki
� �
is computed. The computed d becomes the input of the reversed network, to adjust theweights vector, w. This is essentially the BPN. Thus, we compute for each node uj:
dkj ¼ f 0j
Xiwji · oki
� �t
Xdktw
k21tj
and the weight update is then computed for each wkji [ w:
wkji ¼ wk21
tj þ mdkjoki
The MFN with BPN is known to be far slower and a new learning algorithm based onthe SLFNs, the ELM has been proposed and found to outperform the traditional NN,
Figure 5.A generic model
of neural network
x2
w1
w2
yf(e)
x1
x2
w1Summingjunction
Source: Rumelhart et al. (1986)
e = x1w1+x2w2 y = f(e)Non-linearelement
f(e)w2
x1
ELM for userlocation
prediction
169
and when compared with k-Nearest Neighbour previously used on the same dataset ofSS and SQ (Mantoro et al., 2009), it also presents a better performance. A briefdescription of ELM is given in the subsequent section.
3. Extreme learning machinesThe ELM modeling scheme, a new framework based on the traditional FFNN proposedrecently (Huang et al., 2004), is composed of a single-hidden-layer. It randomly choosesthe input weights and analytically determines the output weights of the SLFNs (Han andHuang, 2006). Previous attempts have compared its performance with other variants ofthe traditional NN as well as other algorithms such as support vector machine, fuzzyneural network, fuzzy regression, multiple linear regression, generalized regressionneural network and other in different domains (Kwak and Kwon, 2008; Wessels andWang, 2010; Huang et al., 2004; Olatunji et al., 2010).
Previous study of the learning rate of FFNN have shown that it is time-consumingwhich hinders FFNN scalability, thereby necessitating the need for new frameworks toformulate faster and better performing networks (Huang et al., 2006a, b). There are twomain reasons behind this behavior (Huang et al., 2009), one is the slow gradient-basedlearning algorithms used to train NN and the other is the iterative tuning of theparameters of the networks by these learning algorithms.
To overcome these problems, Huang et al. (2004) proposed a learning algorithmcalled ELM for SLFNs which randomly selected the input weights and analyticallydetermines the output weights of SLFNs. In Huang et al. (2006a, b), it is stated that“In theory, this algorithm tends to provide the best generalization performance atextremely fast learning speed”. This is a breakthrough since, previously there existed astiff virtual speed barrier which classic learning algorithms could not overcome and,therefore, implementation of FFNN training itself requires longer time, irrespective ofthe application type, whether simple or complex.
The ELM has several interesting and significant features different from traditionalpopular gradient-based learning algorithms for FFNN. These features are enumeratedas follows:
. The learning speed of ELM is extremely fast. In simulations reported previously(Huang et al., 2006a, b, Olatunji et al., 2010), the learning phase of ELM can becompleted in seconds or less than seconds for many applications.
. The ELM has better generalization performance than the gradient-basedlearning such as back propagation in most cases.
. The traditional classic gradient-based learning algorithms and some otherlearning algorithms may face several issues like local minima, improper learningrate, over fitting, etc. In order to avoid these issues, some methods such as weightdecay and early stopping methods may need to be used often in these classicallearning algorithms. The ELM tends to reach the solutions directly without suchtrivial issues. The ELM learning algorithm looks much simpler than mostlearning algorithms for FFNN.
Unlike the traditional classic gradient-based learning algorithms which only work fordifferentiable activation functions, the ELM learning algorithm could be used to trainSLFNs with many non-differentiable activation functions (Huang et al., 2006a, b;Mao et al., 2006).
IJPCC7,2
170
3.1 How extreme learning machine algorithm worksTo appreciate ELM, a description of the standard SLFNs is necessary. A generalframework of extreme learning algorithm is presented here.
Given an N samples (xi, ti), where xi ¼ [xi1, xi2, . . . , xin]T [ Rn andti¼ [ti1, ti2, . . . ,tim]T [ Rn,
The standard SLFNs with ~N hidden neurons and activation function g(x) is defined as:
X~N
i¼1
bigðwi · xj þ biÞ ¼ oj; j ¼ 1; . . . ;N ;
wherewi ¼ [wi1,wi2, . . . ,win]T gives the weight vector that connects the ith hidden neuron
and the input neurons, bi ¼ [bi1,bi2, . . . ,bim]T is the weight vector that connects the ithneuron and the output neurons, and bi is the threshold of the ith hidden neuron. The “· ”in wi · xj indicates the inner product of wi and xj.
In SLFNs, just like the traditional FFNN with BPN, the aim is to minimize thedifference between, network response, oj and the target output, tj. That is:
X~N
i¼1
bigðwi · xj þ biÞ ¼ tj; j ¼ 1; . . . ;N
Or, more compactly, as:
Hb ¼ T
where H can be represented as:
H ðw1; . . . ;w ~N; b1; . . . ; b ~N; x1; . . . ; xN Þ ¼
gðw1 · x1 þ b1Þ · · · gðw ~N · x ~N þ b ~NÞ
..
.· · · ..
.
gðw1 · xN þ b1Þ · · · gðw ~N · xN þ b ~NÞ
26664
37775N¼ ~N
;
b ¼
bT1
..
.
bT~N
266664
377775
~N£m
; and T ¼
TT1
..
.
TT~N
266664
377775N£m
H is referred to as the neural network output matrix (Huang and Haroon, 1998).The processes of the ELM algorithm can be described as follows (Huang et al., 2004):Given a training set:
N ¼ ðxi; tiÞjxi [ R n; ti [ Rm; i ¼ 1; . . . ;Nf g;
activation function g(x), and hidden neuron number ¼ ~N, do the following:. Assign random value to the input weight wi and the bias bi, i ¼ 1, . . . , ~N.. Find the hidden layer output matrix H.. Find the output weight b as follows:
ELM for userlocation
prediction
171
b ¼ H þT
where b, H and T are defined in the same way they were defined in the SLFNsspecification above.
4. Experiment4.1 Experimental dataThe experimental location, Levels 4 and 5 of Kulliyyah of Information andCommunication Technology (KICT) Buildings, International Islamic UniversityMalaysia (IIUM) has wireless infrastructure available through installations ofwireless APs. The locations chosen for the study have a total of 179 room locations,numbered L4001-L4090 and L5001-L5080. The levels have measurable reception ofWi-Fi signals from 8 APs, where each access point produces four service set identifierand media access control addresses on signal reception.
The Wi-Fi SS and SQ were measured and collected from locations, at a distance of1 m in each of the room locations on the chosen levels of the building. In order tofacilitate instance-based learning, the Wi-Fi’s signal data collection was presented in amultivariate signal data.
The data were collected using multi-observers method by collecting the data in aparticular location, a number of times, specifically eight times for all experimentallocations. The calibration was such that the distance between subsequentmeasurements is 1 m.
We collected values of multiple-observed SS and quality in a database, such thatthe ELM algorithm can then estimate the SS and SQ as a data instance from a singlepoint/location which is then compared to the training dataset stored in our database,thereby able to determine the distance value between training data and data sample.The location positioning model is as shown in Figure 6.
Subsequently, the ELM algorithm deduces the closest member based on thetraining dataset by classifying the closest member based on the specified algorithmnumber of closeness. This results in the most user location that belongs to the member(Figure 7).
4.2 Collecting datasetMeasuring and collecting the training dataset are made at Levels 4 and 5 KICTbuilding’s corridor. Tape meters are used to calibrate the floors to measure signal dataat every 1 m. For Level 4, the measurement extends up to 105 m, while Level 5 was upto 88 m.
The data collection was in four folds:First, 8M multi-observer with SS and SQ measurements, by collecting eight data
readings at every single point along the corridor of 90 m at Level 4 and 88 m at Level 5.This results in 1,424 training dataset.
The second stage was collection of 8M multi-observer with SS measurements only,collecting eight data readings at every single point, along the corridor of 90 m at Level4 and 88 m at Level 5, which resulted in 1,424 training dataset.
Next, collection of 1 m multi-observer with SS and SQ measurements. That is,collecting one data reading (SS þ SQ) for both Levels 4 and 5 (Table I).
Finally, 1M multi-observer with SS measurements collecting one data reading (SS)for both Levels 4 and 5.
IJPCC7,2
172
The data were then prepared and stored in a common data repository for the models(Plate 1).
The ELM algorithm was built in the following stages:
(1) The multi-observers training data were partitioned into training and testingdata.
(2) Data preparation was carried out so that the data are well formatted andstructured to conform to the requirements of the network.
Figure 6.The location positioning
architecture
Location
Knowledgebase
Signal converter
Data partitioning
ELM traning
ELM testing
Signal capture
Signal attributes
Figure 7.Experimental locationlayout – user’s within
different location wouldget different signal
strength, quality and noisefrom the wireless access
point
AP1
AP1:–35
AP2:–57
AP3:–78
AP1:–80
AP2:–64
AP3:–39
User1User2
AP2 AP3
ELM for userlocation
prediction
173
(3) The network was trained with different parameters, 80 per cent training datawith 20 per cent test data and 85 per cent training data and 15 per cent test datato compare the performances of each model and, thereby able to select the bettermodel after validation.
An experimental procedure based on test set cross validation was employed in ourstudy. We used the stratified sampling approach to divide the dataset into bothtraining and testing data, such that the size of the training set is 80 per cent of theavailable data and the testing is the rest in the first case, and 85 per cent training datain the second case.
4.3 Simulation settingIn this work, we carried out three separate forms of classification on the same dataset:
(1) classification using the first eight predictor variables (tagged SS1-SS8);
(2) classification using the last eight predictor variables (tagged SQ1-SQ8); and
(3) classification using all the 16 predictor variables (tagged SS1-SS8 andSQ1-SQ8).
For each of the three different cases, the dataset were divided into both training andtesting set using the stratifying approach so that the data were randomly divided foreffective representation. In all, six experiments were recorded as adequaterepresentation among the different models.
4.4 Performance indicatorPercentage of correctly classified samples is the chosen performance indicator for thiswork. These criteria actually provide an easy, clear and more accurate evaluation of theclassifiers. To select the most efficient network based on RMSE and network simplicity,requiring less computation and memory, in the experiment, we differentiated based ononly SS comparing with the results of only SQ and later compare the results obtained basedon these models with the one that used both predictors, SS and SQ combined as dataset.
Plate 1.Capturing the WiFi data
IJPCC7,2
174
SS
SQ
Loc
atio
ns
SS
1S
S2
SS
3S
S4
SS
5S
S6
SS
7S
S8
SQ
1S
Q2
SQ
3S
Q4
SQ
5S
Q6
SQ
7S
Q8
L
272
251
279
299
299
299
299
299
5983
490
00
00
L40
012
692
522
802
992
992
992
992
9963
8248
00
00
0L
4002
264
299
277
299
299
299
299
299
690
520
00
00
L40
032
702
612
762
992
992
992
992
9962
7354
00
00
0L
4004
271
255
299
299
299
299
299
299
6179
00
00
00
L40
052
762
592
762
992
992
992
992
9954
7554
00
00
0L
4006
276
259
299
299
299
299
299
299
5475
00
00
00
L40
072
762
992
992
992
992
992
992
9954
00
00
00
0L
4008
277
252
275
299
299
299
299
299
5282
550
00
00
L40
092
792
492
732
992
992
992
992
9949
8458
00
00
0L
4010
274
251
277
299
299
299
299
299
5783
520
00
00
L40
11
Table I.Sample location data
ELM for userlocation
prediction
175
5. Results and discussionWe present here the results of the classification using 85 per cent for training and15 per cent for testing, then followed by the classification using 80 per cent for trainingand 20 per cent for testing. From the results in Tables II and III, it was shown that theproposed model outperforms the model based on k-Nearest Neighbor obtainedpreviously in Mantoro et al. (2009). The best result from all the parameter variationswas 93.26 per cent in the k-Nearest Neighbor model using 80 per cent training and20 per cent test datasets.
In this model, using all predictors, both SS and SQ, the model gives above 94 per centfor training in both parameter variations of 80-20 per cent training and test data,respectively, and 85-15 per cent variation. Using the latter variation of 85 per centtraining data to 15 per cent for test data was shown to have given the best result,overall, a little above 95 per cent. Generally for the training data, 80 per cent trainingdata for SS alone, without the SQ (SS1-SS8) gives the best performance of slightlyabove 95 per cent. This further confirms that the SS attribute is more relevant in theprediction of location than SQ. This can also be seen in the 85-15 per cent, where the SSalso outperforms the results from SQ. However, the effect of combining both the SQand SS as predictor variables is obvious as we can see that the highest performanceaccuracy was achieved for both training and testing set; specifically, it achievedperformance of above 94 per cent for training set and above 95 per cent for testing set.It can be said also that the average performance for both training and test data isaround 94 per cent using all predictors. This is a more remarkable performance thanthe previous work (Mantoro et al., 2009; Figure 8).
6. ConclusionIn this study, the following conclusions and recommendations could be drawn basedon previous analysis, discussions, deep investigation, experiments and comparativestudies in the work. A new computational intelligence modeling scheme, based on theELM has been investigated, developed and implemented, as an efficient and moreaccurate predictive solution for determining position of mobile users based on locationfingerprinting data – SS and SQ. The new framework based on ELM has beencompared with the k-Nearest Neighbor presented earlier in Mantoro et al. (2009).
SS1-SS8 predictors SQ1-SQ8 predictorsAll predictors
variableTr Ts Tr Ts Tr Ts
ELM 94.72 92.49 93.89 92.96 94.22 95.31
Table II.Results of using85 per cent for trainingand 15 per cent for testing
SS1-SS8 predictors SQ1-SQ8 predictorsAll predictors
variableTr Ts Tr Ts Tr Ts
ELM 95.09 91.55 93.95 93.31 94.47 93.66
Table III.Results of using80 per cent for trainingand 20 per cent for testing
IJPCC7,2
176
The empirical results have shown that the proposed model based on the extremelearning algorithm outperforms k-Nearest Neighbor approaches (Tables II and III).
It is noteworthy to mention based on this work that.While error could lead k-Nearest Neighbor to converge to a wrong prediction, or to
results to multiple or conflicting locations determination on the corridors, because ofdifficulty in conveniently differentiating patterns for each meter, the ELM model wasspecifically invented to overcome lower minima convergence problem, however,a 100 per cent prediction accuracy is still not achieved.
The ELM model outperforms k-Nearest Neighbor (94.22 and 93.26 per cent,respectively) with some noticeable margin.
WiFi’s SS is more relevant to the prediction of location than WiFi’s SQ.Further work is underway to compare the proposed model with other variants of
NN such as BPN, functional networks, simulated, annealing based and further improvethe accuracy with heavy statistics such as Bayesian, Kalman and Particle filters. Wehope to standardize the work to produce reliable campus radio map using a robustlight-weight user location algorithm for mobile devices.
References
Bernacki, M. and Wlodarczyk, P. (2005), “Neural networks and artificial intelligence inelectronics”, available at: www.cs.unibo.it/,roffilli/pub/Backpropagation.pdf (accessed31 August 2010).
Figure 8.Interface for user location
ELM for userlocation
prediction
177
Brimicombe, A. and Li, C. (2009), Location-based Services and Geo-information Engineering,Wiley, Chichester.
Deitel, H.M., Deitel, P.J., Nieto, T.R. and Steinbuhler, K. (2001), Wireless Internet and MobileBusiness How to Program, Prentice-Hall, Upper Saddle River, NJ.
Fox, V., Hightower, J., Liao, L., Schulz, D. and Borriello, G. (2003), “Bayesian filtering for locationestimation”, IEEE Journal of Pervasive Computing, Vol. 2 No. 3, pp. 24-33.
Han, F. and Huang, D. (2006), “Improved extreme learning machine for functionapproximation by encoding a priori information”, Elsevier Journal of Neurocomputing,Vol. 69, pp. 2369-73.
Honkavirta, V., Perala, T., Ali-Loytty, S. and Piche, R. (2009), “A comparative survey of WLANlocation fingerprinting methods”, IEEE Proceedings of the 6th Workshop onPositioning, Navigation and Communication 2009 (WPNC’09), Hannover, Germany,19 March, pp. 243-51.
Huang, G.-B. and Haroon, A.B. (1998), “Upper bounds on the number of hidden neurons infeedforward networks with arbitrary bounded nonlinear activation functions”,IEEE Transactions on Neural Networks, Vol. 9 No. 1, pp. 224-9.
Huang, G.-B., Chen, Y.-Q. and Babri, H.A. (2000), “Classification ability of single hidden layerfeedforward neural networks”, IEEE Transactions on Neural Networks, Vol. 11 No. 3,pp. 799-801.
Huang, G.-B., Zhu, Q.Y. and Siew, C.K. (2004), “Extreme learning machine: a new learningscheme of feedforward neural networks”, Proceedings of the IEEE International JointConference on Neural Networks (IJCNN 2004), Budapest, Hungary, 25-29 July.
Huang, G.B., Zhu, Q.Y. and Siew, C.K. (2006a), “Can threshold networks be trained directly?”,IEEE Transactions on Circuits Systems II, Vol. 53 No. 3, pp. 187-91.
Huang, G.-B., Zhu, Q.Y and Siew, C.K. (2006b), “Extreme learning machine: theory andapplications”, Neurocomputing, Vol. 70 No. 3, pp. 489-501.
Huang, M., Chen, S. and Wang, Y. (2010), “Minimum cost localization problem in wireless sensornetworks publication”, Proceedings of the 7th Annual IEEE Communications SocietyConference on Sensor, Mesh, and Ad Hoc Communications and Networks (SECON 2010),Boston, MA, 21-25 June.
Kaemarungsi, K. (2006), “Distribution of WLAN received signal strength indication for indoorlocation determination”, Technical Report, National Electronics and ComputerTechnology Center, Bangkok.
Kaemarungsi, K. and Krishnamurthy, P. (2004), “Modeling of indoor positioning systems basedon location fingerprinting”, Proceedings of the 23rd IEEE International Conference onComputer Communication (IEEE INFOCOM 2004), Hong Kong, China, 7-11 March.
Kumar, R.K., Apte, V. and Powar, Y.A. (2006), “Improving the accuracy of wireless LAN basedlocation determination systems using Kalman filter and multiple observers”, Proceedingsof the IEEE Wireless Communications and Networking Conferencing (WCNC 2006),New Orleans, LA, pp. 463-8.
Kwak, C. and Kwon, O. (2008), “Cardiac disorder classification based on extreme learningmachine”, Proceedings of World Academy of Science, Engineering and Technology, Rome,Italy, 25-27April, p. 48.
Mantoro, T. and Johnson, C. (2003), “User mobility in an active office”, Proceedings of the FirstEuropean Symposium on Ambient Intelligence, (EUSAI), Veldhoven, The Netherlands,3-4 November.
IJPCC7,2
178
Mantoro, T. and Olowolayemo, A. (2009), “Mobile user location prediction enabled services inubiquitous computing: determining user location using MFN algorithm”, Proceedings ofthe 11th International Conference on Information Integration and Web-based Applications
and Services (iiWAS), Kuala Lumpur, Malaysia, 14-16 December.
Mantoro, T., Azizan, A., Khairuzzaman, S. and Ayu, M.A. (2009), “Multi-observersinstance-based learning approach for indoor symbolic user location determination usingIEEE 802.11 signals”, IEEE Symposium on Industrial Electronics and Applications
(ISIEA 2009), Kuala Lumpur, Malaysia.
Mao, Y., Xuqian, F. and Xue, L. (2006), “A class of self-stabilizing MCA learning algorithms”,IEEE Transactions on Neural Networks, Vol. 17 No. 6, pp. 1634-8.
Mautz, R. (2008), “Combination of indoor and outdoor positioning”, Proceeding of the
1st International Conference on Machine Control, London, UK.
Nuno-Barrau, G. and Paez-Borrallo, J.M. (2005), “A new location estimation system for wirelessnetworks based on linear discriminant functions and hidden Markov models”, EURASIPJournal on Applied Signal Processing, Vol. 2006 No. 1, p. 17 (Article ID 68154).
Olatunji, S.O., Ali, S. and Abdul Azeez, A.-R. (2010), “Modeling permeability prediction usingextreme learning machines”, Proceeding of the 4th Asia International Conference on
Mathematical/Analytical Modelling and Computer Simulation (AMS2010), Kota-Kinabalu,Malaysia, 26-28 May.
Reyero, L. and Delisle, G. (2008), “A pervasive indoor-outdoor positioning system”, Journal ofNetworks, Vol. 3 No. 8, pp. 70-83.
Rohrig, C. and Kunemund, F. (2007), “Estimation of position and orientation of mobile systems ina wireless LAN”, Proceedings of the 46th IEEE International Conference on Decision and
Control, New Orleans, LA, 12-14 December.
Rumelhart, D.E., Hinton, G.E. and Williams, R.J. (1986), “Learning internal representations by errorpropagation”, in Rumelhart, D.E. and McClelland, J.L. (Eds), Parallel Distributed Processing:
Explorations in the Microstructure of Cognition, Vol. 1, MIT Press, Cambridge, MA.
Stanton, N.A. (2001), “Ubiquitous computing: anytime, anyplace, anywhere?”, International
Journal of Journal of Human-computer Interaction, Vol. 13 No. 2, pp. 107-11 ( A specialissue of the international journal of human-computer interaction).
Tan, P., Steinbach, M. and Kumar, V. (2006), Introduction to Data Mining, Pearson International,Boston, MA.
Tsai, C.Y., Chou, S.Y., Lin, S.W. and Wang, W.H. (2009), “Location determination of mobiledevice for indoor WLAN application using neural networks”, Journal for Knowledge andInformation Systems, Vol. 20 No. 1, pp. 81-93.
Wallbaum, M. and Spaniol, O. (2006), “Indoor positioning usingwireless local area networks,jva”, Proceedings of the IEEE John Vincent Atanasoff 2006 International Symposium on
Modern Computing ( JVA’06), Washington, DC, pp. 17-26.
Weiser, M. (1991), “The computer of the twenty-first century”, Scientific American, Vol. 1496,pp. 94-100.
Wessels, A., Wang, X., Laur, R. and Langa, W. (2010), “Dynamic indoor localization usingmultilateration with RSSI in wireless sensor networks for transport logistics”, Proceedingsof Eurosensors XXIV Conference in Procedia Engineering, Linz, Austria, 5-8 September,pp. 220-3.
ELM for userlocation
prediction
179
Further reading
Laurendeau, C. and Barbeau, M. (2010), “Centroid localization of uncooperative nodes in wirelessnetworks using a relative span weighting method”, EURASIP Journal on WirelessCommunications and Networking, Vol. 2010, pp. 1-10 (Article ID 567040).
Roos, T., Myllymaki, P., Tirri, H., Misikangas, P. and Sievanen, J. (2002), “A probabilisticapproach to WLAN user location estimation”, International Journal of WirelessInformation Networks, Vol. 9 No. 3, pp. 155-64.
Yun, S., Lee, J., Chung, W. and Kim, E. (2008), “Centroid localization method in wireless sensornetworks using TSK fuzzy modeling”, Proceedings of the International Symposium OnAdvanced Intelligent Systems, Nagoya, Japan, 17-21 September, pp. 971-4.
Yun, S., Lee, J., Chung, W., Kim, E. and Kim, S. (2009), “A soft computing approach to localizationin wireless sensor networks”, Elsevier Journal of Expert Systems with Applications, Vol. 36,pp. 7552-61.
Zeimpekis, V., Giaglis, G.M. and Lekakos, G. (2003), “A taxonomy of indoor and outdoorpositioning techniques for mobile location services”, ACM SIGecom Exchanges – MobileCommerce, Vol. 3 No. 4, pp. 19-27.
About the authorsTeddy Mantoro (PhD) is currently the Head of Intelligent Environment Research Group (INTEG)and a Senior Lecturer at the Department of Computer Science, KICT, IIUM, Kuala Lumpur.He holds a PhD, an MSc and a BSc, all in Computer Science. He was awarded a PhD fromDepartment of Computer Science, the Australian National University (ANU), Canberra, Australiain 2006. He has authored several research papers, a book on Intelligent Environment, severalchapters in a couple of books and has three patents pending to his credit in the area of pervasivecomputing. Teddy Mantoro is the corresponding author and can be contacted at:[email protected]
Akeem Olowolayemo is a Research Assistant at the INTEG, KICT, IIUM. He holds aBSc (Hons) Degree in Computer Science, from Obafemi Awolowo University Ile-Ife,Nigeria (2000), a Masters (MIT) Degree in Information Systems, from IIUM and is currentlypursuing his PhD (IT) at the same university. His research interests include Location Awareness,Machine Learning and Intelligent Environment.
Sunday O. Olatunji received the BSc (Hons) Degree in Computer Science, Ondo StateUniversity Ado Ekiti, Nigeria (1999), MSc Computer Science, University of Ibadan, Nigeria(2003), MS Degree in Information and Computer Science, King Fahd University of Petroleum andMinerals (KFUPM), Saudi Arabia (2008). He is currently pursuing his PhD in Computer Science.He is a member of ACM and IEEE. He has several years of experience as a lecturer of computerscience, and authored several research papers.
Media A. Ayu is currently a Researcher in the INTEG and a Senior Lecturer at theDeptartment of Information Systems, KICT, IIUM, Kuala Lumpur. He was awarded a PhD fromDepartment of Engineering, College of Engineering and Computer Science, the ANU, Canberra,Australia. She has authored several research papers, several book chapters and has three patentspending to her credit in the area of pervasive computing.
Abu Osman Md. Tab is currently a Researcher in the INTEG and a Professor in Mathematicsat the Department of Information Systems, KICT, IIUM, Kuala Lumpur. He was awarded a PhDfrom University of Birmingham, UK and MSc from Iowa State University, Ames, Iowa, USA.He has authored several research papers, several book and book chapters in the area of fuzzylogic and mathematics.
To purchase reprints of this article please e-mail: [email protected] visit our web site for further details: www.emeraldinsight.com/reprints
IJPCC7,2
180