Expert Systems With Applications 62 (2016) 81–90 Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa EEG-based person identification through Binary Flower Pollination Algorithm Douglas Rodrigues b , Gabriel F.A. Silva a , João P. Papa a , ∗, Aparecido N. Marana a , Xin-She Yang c a São Paulo State University, Department of Computing, Bauru, Brazil b Federal University of São Carlos, Department of Computing, São Carlos, Brazil c Middlesex University, School of Science and Technology, London, UK a r t i c l e i n f o Article history: Received 21 August 2015 Revised 3 June 2016 Accepted 4 June 2016 Available online 14 June 2016 Keywords: Meta-heuristic Pattern classification Biometrics Electroencephalogram Optimum-path forest a b s t r a c t Electroencephalogram (EEG) signal presents a great potential for highly secure biometric systems due to its characteristics of universality, uniqueness, and natural robustness to spoofing attacks. EEG signals are measured by sensors placed in various positions of a person’s head (channels). In this work, we address the problem of reducing the number of required sensors while maintaining a comparable performance. We evaluated a binary version of the Flower Pollination Algorithm under different transf er functions to select the best subset of channels that maximizes the accuracy, which is measured by means of the Optimum-Path Forest classifier. The experimental results show the proposed approach can make use of less than a half of the number of sensors while maintaining recognition rates up to 87%, which is crucial towards the effective use of EEG in biometric applications. © 2016 Elsevier Ltd. All rights reserved. 1 o a r N h s T b o t a s a fi r I d p x e a N a g t o t t t f s t s w i w h i h 0 . Introduction In modern life, we constantly make use of passwords to access ur bank accounts, e-mail boxes, and social networks, just to name few. As passwords can be easily circumvented, the use of biomet- ics has been proposed for safe person identification ( Jain, Ross, & andakumar, 2011 ). Over the years, the use of biometric systems as increased, and systems based on several biometric modalities uch as fingerprint, face and iris, have been successfully deployed. his successful and widespread deployment of biometric systems rings on a new challenge: spoofing. Spoofing methods are devel- ped to breach the security of biometric systems so that unau- horized users can gain access to places and/or information (e.g., n artificial finger made from silicone is placed on the fingerprint canner). In this scenario, the EEG (electroencephalogram) signal presents great potential for highly secure biometric-based person identi- cation, due to its characteristics of universality, uniqueness, and obustness to spoofing attacks ( Beijsterveldt & Boomsma, 1994 ). t is well-known the importance of EEG signals in several ar- ∗ Corresponding author. E-mail addresses: douglasrodrigues.dr@gmail.com (D. Ro- rigues), ec.gabrielalvarez@gmail.com (G.F.A. Silva), papa@fc.unesp.br , apa.joaopaulo@gmail.com (J.P. Papa), nilceu@fc.unesp.br (A.N. Marana), .yang@mdx.ac.uk (X.-S. Yang). i t s p ttp://dx.doi.org/10.1016/j.eswa.2016.06.006 957-4174/© 2016 Elsevier Ltd. All rights reserved. as, since one can find a number of works that deal with such source of data ( Guo, Rivero, Dorado, Munteanu, & Pazos, 2011; unes, Coelho, Lima, Papa, & Albuquerque, 2014; Ocak, 2009; Sub- si, 2007 ). In high security environments, EEG sensors can be inte- rated in order to contribute to the robustness of the system, and he person can be continuously authenticated. Although the idea f using EEG as a biometric trait is not new, there are a few works hat address such kind of signal only. One possible explanation for hat is the difficulty in obtaining such signals, and also because he biometric characteristics of the EEG signal may be held only or short periods of time ( Pollock, Schneider, & Lyness, 1991 ). With the emergence of new mobile devices that capture brain ignals driven by the most keenly studies in the brain computer in- erface, the EEG as a biometric trait can now be used in some other cenarios, such as: (i) distance-based education environments, in hich the continuous authentication of a student becomes increas- ngly necessary; (ii) with the increase in life expectancy world- ide, health monitoring systems may become popular along with ome automation and smart homes, thus making the EEG-based dentification very useful in this scenario; (iii) with the popular- zation of biometric systems for the validation of financial transac- ions, mobile EEG sensors become a viable alternative in the future. Basically, an EEG-based biometric approach aims at placing a et of sensors in the person’s head in order to capture the out- ut signals for further feature extraction and analysis using signal http://dx.doi.org/10.1016/j.eswa.2016.06.006 http://www.ScienceDirect.com http://www.elsevier.com/locate/eswa http://crossmark.crossref.org/dialog/?doi=10.1016/j.eswa.2016.06.006&domain=pdf mailto:douglasrodrigues.dr@gmail.com mailto:ec.gabrielalvarez@gmail.com mailto:papa@fc.unesp.br mailto:papa.joaopaulo@gmail.com mailto:nilceu@fc.unesp.br mailto:x.yang@mdx.ac.uk http://dx.doi.org/10.1016/j.eswa.2016.06.006 82 D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 Fig. 1. International 10-10 System standards for sensor positioning. Just for the sake of clarification, sensor T9 is placed close to the left ear, as well as sensor #23 is placed close to the nose. b m t t T n f m s l b a a s b t i c T s t T t F l t c b e t 3 processing techniques. The signal acquisition session is then re- peated over time to make the system more discriminative and ro- bust to errors. In a recent paper, ( Campisi & La Rocca, 2014 ) pre- sented a review on the state-of-the-art of EEG-based automatic recognition systems, as well as an overview of the neurophysiolog- ical basis that constitutes the foundations on which EEG biometric systems can be built. The authors also discussed about the major obstacles towards the deployment of EEG based biometric systems in everyday life. One of the main problems of EEG-based person identification is the acquisition, which may be too invasive to the user. The pro- cess of putting a considerable amount of sensors up on a person’s head might be a bit uncomfortable, and it also requires a previous knowledge by the person in charge of the sensors placement in order to put them in their correct positions. In light of this con- text, some questions may rise: “Is it really necessary to put all these sensors on a persons’ head? If not, can we identify the most relevant channels for person identification and then use a smaller number of sensors in order to measure them?”. These questions motivated our work in modelling the task of channel selection as an evolutionary-based optimization problem. The idea is to propose a wrapper approach composed by an opti- mization technique and a pattern classifier, in which the accuracy of the latter is used to guide the evolutionary agents in the search space looking for the best solutions, i.e., the subset of channels that maximize the accuracy of the classifier in the validation set. Any optimization technique and classifier could be used. In our work, we propose an optimum channel selection by means of a binary constrained version of the recently proposed optimization technique Flower Pollination Algorithm (BFPA) ( Yang, 2012 ), and the Optimum-Path Forest (OPF) ( Papa, Falcão, Albu- querque, & Tavares, 2012; Papa, Falcão, & Suzuki, 2009 ) classifier, which is a supervised pattern recognition technique that has the advantage of providing a faster training phase compared to other state-of-the-art classifiers. This characteristic of fast training is very important in the context of this paper, since a training procedure followed by a classification of a validation set need to be per- formed for each evolutionary agent (sometimes we may have sev- eral of them). Additionally, this version of OPF is parameterless, which is another advantage over other classifiers. The main contributions of this paper are three-fold: (i) to eval- uate a recent binary version of the Flower Pollination Algorithm (BFPA) proposed by Rodrigues, Yang, Souza, and Papa (2015) un- der different transfer functions 1 ; (ii) to model the problem of EEG channel selection as an evolutionary-based optimization task; and (iii) to introduce the OPF classifier for EEG-based biometric person identification. The use of evolutionary optimization algorithms for the EEG channel selection is due to their elegant and simple so- lutions to solve optimization problems, similar to the way nature does. This paper is organized as follows: Section 2 presents a brief theoretical background about EEG, and Section 3 discusses previ- ous works related to this paper. Section 4 presents the proposed approach for person identification using a reduced number of EEG channels, and Section 5 presents a description of the dataset and the experimental setup. Sections 6 and 7 discuss the experiments and conclusions, respectively. 2. The EEG signal The human central nervous system consists of the encephalous (brain), which is inside the cranium, and the spinal cord contained in the spine. The nerve tissue is a complex network formed mostly 1 A transfer function, in this context, aims at mapping a real-valued solution to a binary-valued one. c ( s y millions of nerve cells (glial cells and neurons), whose pri- ary function is the transmission of electrical impulses that run hrough this intrinsic and huge network, thus propagating informa- ion among cells ( Sanei & Chambers, 2007; Tau & Peterson, 2009 ). hese small electrical impulses emitted by the huge amount of eurons create an electric field that can be measured on the sur- ace of the human skull, with the help of sensors or electrodes. The easurement of this complex electrical signal from our nervous ystem is what is known as electroencephalogram (EEG). In the iterature, it is common among authors to directly refer to those rain waves as EEG. The neural activity of the human being begins between the 17th nd 23rd week of gestation. It is believed that, since this stage, nd throughout the life, the signals from the brain activity repre- ent not only the functioning of the brain, but also of the whole ody. Published studies also show that even if a variation in ampli- ude of EEG signals during the development of a normal person ex- sts, over the years, their functional connections remain largely un- hanged ( Gasser, Jennen-Steinmetz, Sroka, Verleger, & Macks, 1988; au & Peterson, 2009 ). Fig. 1 shows an example of a map of sensors located at a per- on’s head. This map describes the head surface locations via rela- ional distances, also called as International 10-10 System ( Jurcak, suzuki, & Dan, 2007; Nuwer et al., 1998 ). The nomenclature of he electrodes is associated to the human brain areas as follows: rontal (F), Central (C), Temporal (T), Parietal (P) and Occipital (O) obes. Electrodes named with two letters refer to a location be- ween areas, for example: CP electrode is in a position between entral and parietal lobes. The sub-index indicates the side of the rain hemisphere (odd numbers are located on the left side and ven numbers on the right side), and the sub-index “z” indicates hat the electrode is located in the main vertical axis. . Related work One of the first studies regarding EEG as a biometric trait was onducted by Poulos, Rangoussi, Chrissikopoulos, and Evangelou 1999) , which described the EEG signal by means of an autoregres- ive (AR) model as the basis for a person identification method. In D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 83 t m w o s u t 1 f a t w r w t T i a n M f a a s d e S f a d t e p c s n s c g 4 i b e 4 e x w t 2 e s 4 r r a F n 4 i F t t p t r a x w L w c a fl � i a x w s i f i u fl a 4 s p m t a n s t ( S heir work, the correct classification rates reached 91% in experi- ents using data obtained from 45 EEG recordings of 75 subjects, ho were at rest and with the eyes closed during the test. An- ther study by Poulos, Rangoussi, and Alexandris (1999) employed pectral features extracted from the EEG signals followed by the se of neural networks as classifiers to identify a person. The au- hors have achieved correct classification rates ranging from 80% to 00%, reaffirming the great potential of using EEG as a biometric eature. Abdullah, Subari, Loong, and Ahmad (2010) implemented practical system that uses four (sometimes fewer) channels and wo types of EEG signals (one with the eyes open and another one ith the eyes closed), which were used in ten male subjects at est in five different sessions conducted over the course of two eeks. The feature extraction was performed using AR models, and he classification was performed using a multilayer neural network. he authors observed classification rates from 70% to 97%, depend- ng on the amount of channels and EEG type. Palaniappan (2004) used the gamma-band spectral power ratio s features and a Multilayer Perceptron Neural Network to recog- ize a person based on the EEG signal. Later on, Palaniappan and andic (2007) proposed to use 61 channels for feature extraction ollowed by classification using Elman Neural Network. Kostílek nd Št’astný (2012) focused on the importance of the repeatability nd the influence of movements during the EEG signal acquisition ession. In their work, an autoregressive model and a Mahalanobis istance-based classifier for person identification were applied to valuate the robustness of the proposed approach. Safont, Salazar, oriano, and Vergara (2012) used a set of classifiers and multiple eatures to perform EEG-based person identification. In their work, ll possible combinations of features and classifiers have been ad- ressed in order to improve the person recognition results. More recently, La Rocca et al. (2014) proposed a novel approach hat fuses spectral coherence-based connectivity between differ- nt brain regions as a possibly viable biometric feature. The pro- osed approach was tested on a dataset of 108 subjects with eyes- losed (EC) and eyes-open (EO) resting state conditions. Their re- ults show that using brain connectivity leads to higher distinctive- ess when compared with the traditional power-spectrum mea- urements, reaching 100% of recognition accuracy in EC and EO onditions when integrating functional connectivity between re- ions in the frontal lobe. . Proposed method In this section, we present our proposed method for person dentification based on features from EEG signals, as well as we riefly review some of the main concepts regarding the techniques mployed in this paper. .1. Autoregressive model An autoregressive model can be described by a linear difference quation in the time domain as follows: (k ) = P + p ∑ i =1 a (i ) x (t − i ) + e (t) , (1) here P is a constant, p stands for the number of parameters of he model and e ( t ) denotes a white noise input ( Jain & Deshpande, 004 ). Notice In this work, we used the Yule–Walker method to stimate the coefficients of the AR model by employing the least quare method criterion. .2. EEG channel selection In order to select the best subset of channels, we evaluate a ecent proposed binary version of the Flower Pollination Algo- ithm ( Rodrigues et al., 2015 ) under different transf er functions, nd we also show we can obtain distinct results for each one. irstly, we present the theoretical basis about FPA, and then its bi- ary version. .2.1. Flower pollination algorithm The Flower Pollination Algorithm proposed by Yang (2012) is nspired by the flow pollination process of flowering plants. The PA is governed by four basic rules: 1. Biotic cross-pollination can be considered as a process of global pollination, and pollen-carrying pollinators move in a way that obeys Lévy flights; 2. For local pollination, abiotic pollination and self-pollination are used; 3. Pollinators such as insects can develop flower constancy, which is equivalent to a reproduction probability that is proportional to the similarity of two flowers involved; and 4. The interaction or switching of local pollination and global pol- lination can be controlled by a switch probability p ∈ [0, 1], slightly biased towards local pollination. In order to model the updating formulas, the above rules have o be converted into proper updating equations. For example, in he global pollination step, flower pollen gametes are carried by ollinators such as insects, and pollen can travel over a long dis- ance because insects can often fly and move over a much longer ange. Therefore, Rules 1 and 3 can be represented mathematically s follows: (t+1) i = x t i + αL (λ)(g ∗ − x t i ) , (2) here (λ) = λ · �(λ) · sin (λ) π · 1 s 1+ λ , s � s 0 > 0 (3) here x t i is the pollen i (solution vector) at iteration t, g ∗ is the urrent best solution among all solutions at the current generation, nd α is a scaling factor to control the step size. L ( λ) is the Lévy- ights step size, that corresponds to the strength of the pollination, ( λ) stands for the gamma function and s is the step size. Since nsects may move over a long distance with various distance steps, Lévy flight can be used to mimic this characteristic efficiently. For local pollination, both Rules 2 and 3 can be represented as: (t+1) i = x t i + ε(x t j − x t k ) , (4) here x t j and x t k are pollen from different flowers j and k of the ame plant species at time step t . This mimics flower constancy n a limited neighbourhood. Mathematically, if x t j and x t k come rom the same species or are selected from the same population, t equivalently becomes a local random walk if ε is drawn from a niform distribution in [0,1]. In order to mimic the local and global ower pollination, a switch probability (Rule 4) or proximity prob- bility p is used. .2.2. Binary flower pollination algorithm In the standard FPA, the solutions are updated in the search pace towards continuous-valued positions. However, in the pro- osed Binary Flower Pollination Algorithm the search space is odelled as an n -dimensional boolean lattice, in which the solu- ions are updated across the corners of a hypercube. In addition, s the problem is to select or not a given feature, a solution bi- ary vector is employed, where 1 corresponds to a feature being elected to compose the new set, and 0 otherwise. In order to build his binary vector, ( Rodrigues et al., 2015 ) employed Eqs. (5) and 6 ), which can restrict the new solutions to only binary values: (x j i (t)) = 1 1 + e −x j i (t) , (5) 84 D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 a p 4 P m b b t t p O l a 5 s s g 5 M d S 2 s s m n a a t o t o w i 6 o c fi t e o h u a 2 http://physionet.org/pn4/eegmmidb . x j i (t) = { 1 if S(x j i (t)) > σ, 0 otherwise (6) in which σ ∼ U (0, 1). Algorithm 1 presents the proposed approach that employs BFPA for EEG-channel selection using the OPF classi- fier as the objective function and Eqs. (5) and ( 6 ) as the transfer function. Note that the proposed approach can be used with any other supervised classification technique. Algorithm 1: BFPA - Binary Flower Pollination Algorithm input : Training set Z 1 and evaluating set Z 2 , α, number of flowers m , dimension d and iterations T . output : Global best position ̂ g . auxiliaries : Fitness vector f with size m and variables acc, max f it , gl obal f it ← −∞ and maxindex . 1 for each flower i (∀ i = 1 , . . . , m ) do 2 for each dimension j (∀ j = 1 , . . . , d) do 3 x j i (0) ← Random { 0 , 1 } ; 4 f i ← −∞ ; 5 for each iteration t (t = 1 , . . . , T ) do 6 for each flower i (∀ i = 1 , . . . , m ) do 7 Create Z ′ 1 and Z ′ 2 from Z 1 and Z 2 , respectively, such that both contains only features such that x j i (t) = 0 , ∀ j = 1 , . . . , d; 8 Train OPF over Z ′ 1 , evaluate its over Z ′ 2 and stores the accuracy in acc; 9 if (acc > f i ) then 10 f i ← acc; 11 for each dimension j (∀ j = 1 , . . . , d) do 12 ̂ x j i ← x j i (t) ; 13 [ max f it, maxindex ] ← max ( f ) ; 14 if (max f it > gl obal f it) then 15 gl obal f it ← max f it; 16 for each dimension j (∀ j = 1 , . . . , d) do 17 ̂ g j ← x j maxindex (t) ; 18 for each flower i (∀ i = 1 , . . . , m ) do 19 for each dimension j (∀ j = 1 , . . . , d) do 20 rand ← Random { 0 , 1 } ; 21 if rand < p then 22 x j i (t) ← x j i (t − 1) + α � Lévy (λ) ; else 23 x j i (t) ← x j i (t − 1) + ε(x j i (t − 1) − x k i (t − 1)) ; 24 if (σ < 1 1+ e x j i (t) ) then 25 x j i (t) ← 1 ; else 26 x j i (t) ← 0 ; Lines 1–4 initialize each pollen’s position as being a binary string with random values, as well as the fitness value f i of each individual i . The main loop in Lines 6–27 is the core of the pro- posed algorithm, in which the inner loop in Lines 7–13 is respon- sible for creating the new training Z ′ 1 and evaluating sets Z ′ 2 , and then OPF is trained over Z ′ 1 and it is used to classify Z ′ 2 . The recog- nition accuracy over Z ′ 2 is stored in acc and then compared with the fitness value f i (accuracy) of individual i : if the later is worse than acc , the old fitness value is kept; in the opposite case, the fit- ness value is then updated. Lines 12–13 update the best local posi- tion of the current pollen. Lines 14–18 update the global optimum, nd the last loop (Lines 19–27) moves each pollen to a new binary osition restricted by Eqs. (5) and ( 6 ) (Lines 25–27). .3. Optimum-path forest classifier We used the Optimum-Path Forest classifier ( Papa et al., 2012; apa et al., 2009 ) applied to the features learned from the AR odel to classify a person based on the EEG signal. The OPF works y modelling the samples as graph nodes, whose arcs are defined y an adjacency relation and weighted by a distance function. Fur- her, a role competition process between some key nodes (proto- ypes) is carried out in order to partition the graph into optimum- ath trees (OPTs) according to a path-cost function. In fact, each PT is rooted at one prototype, which means a sample that be- ongs to a given tree is more strongly connected to its root than to ny other in the forest. . Methodology In this section, we present the proposed approach for channel election in EEG-based signal acquisition, as well as we briefly de- cribe the employed dataset, the nature-inspired meta-heuristic al- orithms, and the experimental setup. .1. Dataset The EEG signals used in this work were obtained from the EEG otor Movement/Imagery dataset 2 ( Goldberger et al., 20 0 0 ). The ata was collected from 109 healthy volunteers using the BCI20 0 0 ystem ( Schalk, McFarland, Hinterberger, Birbaumer, & Wolpaw, 004 ), which makes use of 64 channels (sensors) and provides a eparated EDF (European Data Format) file for each of them. The ubjects performed different motor/imagery tasks: such tasks are ainly used in BCI (Brain-Computer Interface) applications and eurological rehabilitation, and consists of imagining or simulating given action, like open and close the eyes, for example. Each subject performed four tasks according to the position of target that appears on the screen placed in front of the volun- eers (if the target appears on the right or left side, the subject pens and closes the corresponding fist; if the target appears on he top or bottom side, the subject opens and closes both fists r both feets, respectively). In short, the four experimental tasks ere: 1. To open and close left or right fist; 2. To imagine opening and closing left or right fist; 3. To open and close both fists or both feet; and 4. To imagine opening and closing both fists or both feet. Each of these tasks were performed three times, thus generat- ng 12 recordings for each subject of a two-minutes run, and the 4 channels were sampled at 160 samples per second. The features of the twelve recordings are extracted by means f an AR model with three output configurations for each EEG- hannel: 5, 10 and 20 features. Further, the average of each con- guration is then been computed in order to obtain just one fea- ure per EEG-channel (sensor). In short, for each sensor, we have xtracted three different numbers of AR-based features, being the utput of each sensor the average of their values. Henceforth, we ave adopted the following notation for each of the dataset config- rations: AR 5 for 5 autoregression coefficients extracted, and AR 10 nd AR 20 for 10 and 20 autoregression coefficients, respectively. http://physionet.org/pn4/eegmmidb D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 85 5 o m 1 ( ( & m t f t 5 s a o t Table 1 Parameters used for each meta-heuristic optimization tech- nique. Notice the inertia weight w for PSO was linearly de- creased from 0.9 to 0.4. Technique Parameters BGA mutation = 0 . 1 BPSO c 1 = c 2 = 2 BFA γ = 0 . 8 , β0 = 1 . 0 , α = 0 . 01 BCSS – BHS HMCR = 0 . 9 BFPA α = 1 . 0 , p = 0 . 8 t w o t f a i m i N m d e n c M o t m fi 6 s s d e C t f n s F a c n e a p t a d i o 3 We have used the same variable notation for different methods because we be- lieve it makes it easier to understand since it is the same notation used in the respective original papers. .2. Nature-inspired meta-heuristic algorithms In this work, we have compared our proposed method with ther meta-heuristic-based optimization methods described below: Genetic algorithm (GA): The Genetic algorithm was proposed by Holland (1975) , and its main concept is to emulate the bi- ological evolution to solve optimization problems. It is com- posed of an initial population (or a set of unique elements) and a set of operators inspired by the nature. These opera- tors can change the elements, and according to the evolu- tionary theory, only the most capable individuals are able to survive and transmit their biological heredity to the next generations. Particle swarm optimization (PSO): This method is inspired on the social behaviour of a bird flocking or a fish school- ing ( Kennedy & Eberhart, 2001 ). The fundamental idea is that each particle represents a potential solution which is updated according to its own experience and from its neigh- bours’ knowledge. The motion of an individual particle for the optimal solution is governed through its position and ve- locity interactions, and also by its own previous best perfor- mance and the best performance of their neighbours. Firefly algorithm (FA): This method was proposed by Yang (2010) , being derived from the flash attractiveness of fire- flies for mating partners (communication) and attracting po- tential preys. The brightness of a firefly at a given position is determined by the value of the objective function in that position. Each firefly is attracted by a brighter firefly through the attraction factor. Harmony search (HS): This method is a meta-heuristic algo- rithm inspired in the improvisation process of music play- ers ( Geem, 2009 ). Musicians often improvise the pitches of their instruments searching for a perfect state of harmony. The main idea is to use the same process adopted by musi- cians to create new songs to obtain a near-optimal solution according to some fitness function. Each possible solution is modelled as a harmony, and each musical note corresponds to one decision variable. Charged system search (CSS): This method, based on the governing Coulomb’s law (a physics law used to describe the interactions between electrically charged particles), was proposed by Kaveh and Talatahari (2010) . In this method, named CSS, each Charged Particle (CP) in the system is af- fected by the electrical fields of the others, generating a re- sultant force over each CP, which is determined by using the electrostatic laws. The CP interaction movement is de- termined by Newtonian mechanics laws. We have used the binary optimization version of each afore- entioned method, as proposed in: Binary GA (BGA) ( Holland, 975 ), Binary PSO (BPSO) ( Firpi & Goodman, 2004 ), Binary HS BHS) ( Ramos, Souza, Chiachia, Falcão, & Papa, 2011 ), Binary Firefly BFA) ( Falcon, Almeida, & Nayak, 2011; Palit, Sinha, Molla, Khanra, Kule, 2011 ), and Binary CSS ( Rodrigues et al., 2013 ). The opti- ization algorithms were implemented in C language following he guidelines provided by their references. Notice the transfer unction defined by Eqs. (5) and ( 6 ) were the very same for all echniques compared in this work. .3. Experimental setup We partitioned our fully labeled dataset into Z = Z 1 ∪ Z 2 ∪ Z 3 ubsets, in which Z 1 , Z 2 and Z 3 stand for training, validation, nd test sets, respectively. The training dataset contains 50% of the riginal dataset, followed by 30% and 20% concerning the valida- ion and test sets, respectively. The idea is to employ Z and Z 1 2 o find the subset of features that maximize the accuracy over Z 2 , ith the accuracy being the fitness function. Each agent is initialized with random binary positions and the riginal dataset is mapped to a new one that contains the features hat were selected in this first sampling. In addition, the fitness unction of each agent is set to the OPF recognition rate over Z 2 fter training in Z 1 . The final subset will be the one that max- mizes the curve over the range of values, i.e., the features that aximize the accuracy over Z 2 . The accuracy over the test set Z 3 s then assessed by using the final subset of the selected features. otice the fitness function employed in this paper is the accuracy easure proposed by Papa et al. (2009) , which is capable of han- ling unbalanced classes. Fig. 2 presents the methodology used to valuate the proposed approach. Table 1 shows the parameters used for each optimization tech- ique employed in this work 3 . The c 1 and c 2 parameters of PSO ontrol the pace during the particles movement, and the “Harmony emory Considering Rate” (HMCR) of BHS stands for the amount f information that will be used from the artist’s memory (songs hat have been already composed) in order to compose a new har- ony. In regard to BFA, α and β0 are related to the step size of a refly, and γ stands for the light absorption coefficient. . Experimental results The experimental results stand for the mean accuracy and tandard deviation over 25 rounds using the methodology pre- ented in Section 5.3 . Since the meta-heuristic algorithms are non- eterministic, we adopt this protocol to avoid biased results. The xperiments were executed in a computer with a Pentium Intel ore i 7 ® 1.73Ghz processor, 6 GB of RAM and Linux Ubuntu Desk- op LTS 13.04 as the operational system. Figs. 3 and 4 present the mean OPF accuracy over the three dif- erent feature sets (AR 5 , AR 10 and AR 20 ), as well as the average umber of selected channels, respectively. Notice the “yellow” bar tands for the standard OPF, i.e., without channel selection. From ig. 3 , one can observe there is not a relevant difference in terms of ccuracy considering the different number of autoregression coeffi- ients. As the coefficients are averaged at the output of each chan- el, such non-linear operation may have alleviated the influence of ach approach. However, this operation seems to work well, since recognition rate of around 86% is very competitive when com- ared to other works in the literature ( Section 3 ). Table 2 presents the percentage of selected EEG-channels. From he data, it is possible to observe three important points: (i) BGA nd BHS have selected the lowest number of channels for all ataset configurations; (ii) considering the accuracy results shown n Fig. 3 , we can conclude that we can achieve similar performance f that obtained using all the 64 channels by using less than a half 86 D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 Fig. 2. Block diagram of the proposed approach. Fig. 3. Average OPF accuracy over (a) AR 5 , (b) AR 10 and (c) AR 20 configurations. Fig. 4. Average number of selected channels of all techniques over (a) AR 5 , (b) AR 10 and (c) AR 20 configurations. These values have been truncated for sake of simplicity. Table 2 Percentual of selected EEG-channels. Dataset BGA BPSO BFA BHS BCSS BFPA AR 5 36% 38% 45% 38% 44% 46% AR 10 36% 39% 44% 36% 45% 45% AR 20 37% 40% 44% 36% 44% 45% p e e h d t t f ( c t T s of them; and (iii) the proposed BFPA has been very competitive in terms of binary-constrained optimization tasks when compared to the techniques addressed in this work. Fig. 5 depicts the mean computational load (in seconds) for all optimization techniques regarding the learning step (dark gray module in Fig. 2 ). As we did not consider the feature extraction rocedure, i.e., the autoregression coefficients computation, the ex- cution time over all dataset configurations are quite similar for ach specific optimization technique. It is possible to observe BHS as been the fastest technique in all situations, since it only up- ates one agent per iteration. Although it may be a drawback in erms of convergence, it is still the fastest approach. Finally, we performed the Wilcoxon signed-rank statistical est ( Wilcoxon, 1945 ) to verify whether there is a significant dif- erence between BFPA and the other techniques used in this work considering the OPF recognition rate). Table 3 displays a pair-wise omparison against all techniques and BFPA, showing whether two echniques are considered similar (‘ = ’) or not (‘ = ’) to each other. he only technique that has been considered similar to BFPA in all ituations is BFA, followed by BPSO. An interesting point is related D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 87 Fig. 5. Mean execution times of all techniques over (a) AR 5 , (b) AR 10 and (c) AR 20 configurations. Fig. 6. Frequency of selected sensors during the experimental evaluation using AR 5 and BFPA. Table 3 Wilcoxon signed-rank test evaluation. Dataset BGA BPSO BFA BHS BCSS AR 5 = = = = = AR 10 = = = = = AR 20 = = = = = t w s m t W e t v b f a t i b t s b r l i s [ t o 6 s t 2 f b o o t w t a e I F t M c e i L i v g t v c g S o the number of parameters, since BFPA requires only two, mean- hile BFA needs three parameters. Since the nature of the proposed task in the EEG recording ession has a close relation with different brain areas, like the ovements of the hands and feet that mainly activates the cen- ral region of the brain ( Wang, Gao, & Gao, 2005; Yang, Kyrgyzov, iart, & Bloch, 2013 ), it is important to figure out whether the xpected channels are actually included in the subset selected by he optimization techniques. Therefore, since we executed a cross- alidation procedure with 25 runnings, and due to the stochastic ehaviour of the meta-heuristic techniques, this means a certain eature may not be selected at a given execution, and may be at nother. In order to cope with this challenge, we opted to display he frequency of occurrence concerning each sensor, as displayed n Fig. 6 . In this case, we considered BFPA with feature extraction y model AR . 5 Some interesting conclusions can be drawn if we consider he different range of frequencies modelled by distinct colours. It eems the frontal sensors are slightly more important than the ack ones, since we can find more “yellow” and “blue” sensors ight below the horizontal line (i.e., the one that goes from the eft ear to the right one) than above that line. Another observation s that the “yellow” sensors are place everywhere, i.e., they corre- pond to the sensors that have been selected in between the range 85%, 89%], which is a considerable frequency. This means BFPA ried to select sensors placed at different positions of the brain in rder to capture different information. .1. Transfer function analisys In order to map the possible solutions (i.e., a position in the earch space) from a continuous-valued space to a binary one, a ransfer function needs to be employed ( Mirjalili & Mohd Hashim, 011; Rashedi, Nezamabadi-pour, & Saryazdi, 2010 ). A transfer unction defines the probability of changing the position of a possi- le solution from 0 to 1 and vice-versa forcing the agents to move nto a binary space. Mirjalili and Lewis (2013) introduced a study f two families of transfer functions on binary-based PSO. Since he binary version of FPA makes use of a transfer function either, e also investigated these two different families of transfer func- ions (S-shaped and V-shaped) on Binary FPA. In short, we evalu- ted 8 transfer functions, as follows: • S-shaped: S1, S2, S3 and S4; and • V-shaped: V1, V2, V3 and V4. Notice the transfer function S2 is the same one used in the xperiments conducted in the previous section ( Eqs. (5) and ( 6 )). n this section, we just reproduced the results obtained with S2. or a more detailed explanation about the functions employed in his section, the reader can refer to the work by Mirjalili and ohd Hashim (2011) ; Rashedi et al. (2010) . First of all, we evaluated the convergence of all tranfer functions onsidering the AR models used in this work. Fig. 7 displays this xperiment, in which transfer function S1 obtained the best results n all AR models, followed by S2 and V1. According to Mirjalili and ewis (2013) , the larger the velocity of a given particle, the highest t should be the probability to change its position from 1 to 0 and ice-versa, since this particle probably is far away from the best lobal solution. In this context, the “most abrupt” transfer func- ions are S1 and V1, i.e., they are more prone to switch the binary alues. Following a similar behaviour to the ones obtained in the onvergence-driven experiment, functions S1 and V1 provided very ood recognition rates over the test set, as displayed in Fig. 8 . uch behaviour can be observed for all AR models. Additionally, 88 D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 Fig. 7. Convergence evaluation of the transfer functions considering all AR models. Fig. 8. Average OPF accuracy over (a) AR 5 , (b) AR 10 and (c) AR 20 configurations considering different transfer functions. Fig. 9. Average number of selected channels of all techniques over (a) AR 5 , (b) AR 10 and (c) AR 20 configurations considering different transf er functions. These values have been truncated for sake of simplicity. 6 r B m t c s r t o T the number of selected features can influence the recognition rates, as one can observe in Fig. 9 . Although transfer function V3 has selected less features, it obtained the lowest recognition rates ( Fig. 8 ), which is somehow expected. In regard to the computa- tional load, Fig. 10 presents the mean execution time to learn the most representative subset of features. Since transfer function S1 has selected more features, it is expected a higher computational burden when compared to the others. Table 4 displays the Wilcoxon signed-rank test considering the experiment with different transfer functions. Considering model AR 5 , the most accurate techniques were S1, S2 and S4, and with respect to AR 10 we can highlight S1, S4 and V2 as the top-3 tech- niques. Finally, S1 and S4 obtained the best results considering the model AR 20 . .2. Discussion Roughly speaking, all techniques achieved similar recognition ates considering all AR models, with an advantage to BFPA and FA, which are swarm-oriented. It is important to highlight one ight obtain better recognition rates using a different feature ex- raction, but the main goal of this work is to evaluate BFPA in the ontext of sensor selection, as well as to show the importance of electing sensors in order to make such approach less prone to er- ors and probably cheaper. Using AR models with different number of coefficients seemed o does not provide different recognition rates, since the output f each AR model is given by the average of the coefficients. his could be a plausible explanation for that case. Such assump- D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 89 Fig. 10. Mean execution times of all techniques considering different transfer functions over (a) AR 5 , (b) AR 10 and (c) AR 20 configurations. Table 4 Wilcoxon signed-rank test computed between the transfer functions. AR 5 S1 S2 S3 S4 V1 V2 V3 V4 S1 — = = = = = = = S2 = — = = = = = = S3 = = — = = = = = S4 = = = — = = = = V1 = = = = — = = = V2 = = = = = — = = V3 = = = = = = — = V4 = = = = = = = — AR 10 S1 S2 S3 S4 V1 V2 V3 V4 S1 — = = = = = = = S2 = — = = = = = = S3 = = — = = = = = S4 = = = — = = = = V1 = = = = — = = = V2 = = = = = — = = V3 = = = = = = — = V4 = = = = = = = — AR 20 S1 S2 S3 S4 V1 V2 V3 V4 S1 — = = = = = = = S2 = — = = = = = = S3 = = — = = = = = S4 = = = — = = = = V1 = = = = — = = = V2 = = = = = — = = V3 = = = = = = — = V4 = = = = = = = — t p B a a B i h s l s f t s p t 7 b h o m o t f O i n w t t n P m o r r c I h i a n w l l s s P c b p fi A g ion can be applied to all meta-heuristic techniques used in this aper. Another important point concerns with the sensors selected by FPA. A more detailed study showed the most frequent sensors re located in the front of the head, tough they are also spread long the head. That is an interesting observation, which means FPA tried to select sensors that are not so close to each other n order to capture relevant information from all places of the ead. Finally, an additional study with different transf er functions howed we can obtain different results, being the number of se- ected features strongly related to the final recognition rates. It eems the more features one has, the most accurate the transfer unction. However, we still need to deal with a trade-off between he number of features and the computational efficiency. Using all ensors does not give us too much different results, which sup- orts the idea of this work, that is to emphasize one can find out he subset of sensors that can obtain reasonable results. . Conclusions and future work We have addressed the problem of channel selection in EEG- ased biometric person identification. The goal of this work to ighlight we may not need to employ all EEG channels available in rder to obtain high identification rates. Therefore, we proposed to odel the problem of channel selection as a meta-heuristic-based ptimization task, in which the subset of channels that maximize he recognition rate over a validation set is used as the fitness unction. For the identification (classification) task, we have used the ptimum-Path Forest classifier, which has demonstrated to be sim- lar to the state-of-the-art supervised pattern recognition tech- iques, but faster for training. In regard to the meta-heuristics, e have introduced a binary-constrained optimization version of he recently proposed Flower Pollination Algorithm, which seemed o be very competitive to other state-of-the-art optimization tech- iques employed in this paper: Binary Genetic Algorithm, Binary article Swarm Optimization, Binary Firefly Algorithm, Binary Har- ony Search, and Binary Charged System Search. The experimental results showed the BFPA outperformed many f the other methods, obtaining very good person identification ates using much less channels. It is important to emphasize that educing EEG channels while keeping high identification rates is rucial towards the effective use of EEG in biometric applications. n addition, the selected sensors seemed to cover all the person’s ead, mainly in the front. Moreover, the number of coefficients n the AR model does not seem to impact in the final results, lthough we are taking the average of the coefficients as the fi- al feature. Finally, different transfer functions were also analyzed, hich allowed slightly better results. Although using EEG data for biometric purposes seems to be a ittle bit far from reality in non-controlled environments, we would ike to shed light over the importance in keep going with such tudies, since good recognition rates can be obtained, being such ort of biometric approaches much less prone to spoofing attacks. robably, in the future when mobile devices can be used to easily apture EEG signals, such techniques can be widely employed for iometric purposes as well. Our future work will involve using modified versions of FPA to erform channel selection aiming at improving the overall identi- cation performance while selecting fewer channels. cknowledgments The authors are grateful to FAPESP grant #2014/16250-9, CNPq rants #470571/2013-6 and #30 616 6/2014-3, and Capes grant. 90 D. Rodrigues et al. / Expert Systems With Applications 62 (2016) 81–90 P P P P P R S S W W Y Y Y References Abdullah, M. K. , Subari, K. S. , Loong, J. L. C. , & Ahmad, N. N. (2010). Analysis of effective channel placement for an eeg-based biometric system. In Proceedings of the IEEE conference on biomedical engineering and sciences (pp. 303–306) . Beijsterveldt, C. , & Boomsma, D. (1994). Genetics of the human electroencephalo- gram (EEG) and event-related brain potentials (ERPs): A review. Human Genetics, 94 (4), 319–330 . Campisi, P., & La Rocca, D. (2014). Brain waves for automatic biometric-based user recognition. Information Forensics and Security, IEEE Transactions on, 9 (5), 782– 800. doi: 10.1109/TIFS.2014.2308640 . Falcon, R. , Almeida, M. , & Nayak, A. (2011). Fault identification with binary adaptive fireflies in parallel and distributed systems. In Evolutionary computation (CEC), 2011 IEEE congress on (pp. 1359–1366) . Firpi, H. A. , & Goodman, E. (2004). Swarmed feature selection. In Proceedings of the 33rd applied imagery pattern recognition workshop (pp. 112–118). Washington, DC, USA: IEEE Computer Society . Gasser, T. , Jennen-Steinmetz, C. , Sroka, L. , Verleger, R. , & Macks, J. (1988). Develop- ment of the eeg of school-age children and adolescents ii. topography. Electroen- cephalography and clinical neurophysiology, 69 (2), 100–109 . Geem, Z. W. (2009). Music-inspired harmony search algorithm: Theory and applica- tions (1st ed.). Springer Publishing Company, Incorporated . Goldberger, A. L. , Amaral, L. A. N. , Glass, L. , Hausdorff, J. M. , Ivanov, P. C. , Mark, R. G. , et al. (20 0 0). Physiobank, physiotoolkit, and physionet: Components of a new research resource for complex physiologic signals. Circulation, 101 (23), e215–e220 . Guo, L. , Rivero, D. , Dorado, J. , Munteanu, C. R. , & Pazos, A. (2011). Automatic feature extraction using genetic programming: An application to epileptic EEG classifi- cation. Expert Systems with Applications, 38 (8), 10425–10436 . Holland, J. H. (1975). Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control and artificial intelligence. . Oxford, England: U Michigan Press . Jain, A. K. , Ross, A. A. , & Nandakumar, K. (2011). Introduction to biometrics . Springer . Jain, S. , & Deshpande, G. (2004). Parametric modeling of brain signals. In Biotech- nology and bioinformatics, 2004. proceedings. technology for life: North carolina symposium on (pp. 85–91) . Jurcak, V. , Tsuzuki, D. , & Dan, I. (2007). 10/20, 10/10, and 10/5 systems revisited: Their validity as relative head-surface-based positioning systems. NeuroImage, 34 (4), 1600–1611 . Kaveh, A. , & Talatahari, S. (2010). A novel heuristic optimization method: charged system search. Acta Mechanica, 213 (3), 267–289 . Kennedy, J. , & Eberhart, R. C. (2001). Swarm Intelligence . M. Kaufman . Kostílek, M. , & Št’astný, J. (2012). EEG biometric identification: Repeatability and influence of movement-related EEG. In Proceedings of the international conference on applied electronics (pp. 147–150) . La Rocca, D., Campisi, P., Vegso, B., Cserti, P., Kozmann, G., Babiloni, F., et al. (2014). Human brain distinctiveness based on EEG spectral coherence connec- tivity. Biomedical Engineering, IEEE Transactions on, 61 (9), 2406–2412. doi: 10. 1109/TBME.2014.2317881 . Mirjalili, S. , & Lewis, A. (2013). S-shaped versus v-shaped transfer functions for bi- nary particle swarm optimization. Swarm and Evolutionary Computation, 9 , 1–14 . Mirjalili, S. , & Mohd Hashim, S. Z. (2011). BMOA: Binary magnetic optimization al- gorithm. In 3rd IEEE international conference on machine learning and computing: vol. 1 (pp. 201–206) . Singapore. Nunes, T. M. , Coelho, A. L. , Lima, C. A. M. , Papa, J. P. , & Albuquerque, V. H. C. (2014). EEG signal classification for epilepsy diagnosis via optimum path forest – a sys- tematic assessment. Neurocomputing, 136 , 103–123 . Nuwer, M. R. , Comi, G. , Emerson, R. , Fuglsang-Frederiksen, A. , Guérit, J. M. , Hin- richs, H. , et al. (1998). IFCN standards for digital recording of clinical EEG. Elec- troencephalography and Clinical Neurophysiology, 106 (3), 259–261 . Ocak, H. (2009). Automatic detection of epileptic seizures in EEG using discrete wavelet transform and approximate entropy. Expert Systems with Applications, 36 (2, Part 1), 2027–2036 . Palaniappan, R. (2004). Method of identifying individuals using VEP signals and neural network. IEE Proceedings - Science, Measurement and Technology, 151 (1), 16–20 . alaniappan, R. , & Mandic, D. (2007). EEG based biometric framework for automatic identity verification. The Journal of VLSI Signal Processing Systems for Signal, Im- age, and Video Technology, 49 (2), 243–250 . alit, S. , Sinha, S. N. , Molla, M. A. , Khanra, A. , & Kule, M. (2011). A cryptanalytic attack on the knapsack cryptosystem using binary firefly algorithm. In Com- puter and communication technology (ICCCT), 2011 2nd international conference on (pp. 428–432) . Papa, J. P. , Falcão, A. X. , Albuquerque, V. H. C. , & Tavares, J. M. R. S. (2012). Efficient supervised optimum-path forest classification for large datasets. Pattern Recog- nition, 45 (1), 512–520 . apa, J. P. , Falcão, A. X. , & Suzuki, C. T. N. (2009). Supervised pattern classification based on optimum-path forest. Intl. Journal of Imaging Systems and Tech., 19 (2), 120–131 . ollock, V. , Schneider, L. , & Lyness, S. (1991). Reliability of topographic quantita- tive EEG amplitude in healthy late-middle-aged and elderly subjects. Electroen- cephalography and Clinical Neurophysiology, 79 (1), 20–26 . Poulos, M. , Rangoussi, M. , & Alexandris, N. (1999). Neural network based person identification using EEG features. In Proceedings of the IEEE international confer- ence on acoustics, speech, and signal processing (pp. 1117–1120) . oulos, M. , Rangoussi, M. , Chrissikopoulos, V. , & Evangelou, A. (1999). Parametric person identification from the EEG using computational geometry. In Proceed- ings of IEEE international conference on electronics, circuits and systems, 1999: vol. 2 (pp. 1005–1008) . Ramos, C. C. O. , Souza, A. N. , Chiachia, G. , Falcão, A. X. , & Papa, J. P. (2011). A novel algorithm for feature selection using harmony search and its applica- tion for non-technical losses detection. Computers & Electrical Engineering, 37 (6), 886–894 . ashedi, E. , Nezamabadi-pour, H. , & Saryazdi, S. (2010). BGSA: Binary gravitational search algorithm. Natural Computing, 9 (3), 727–745 . Rodrigues, D. , Pereira, L. A. M. , Papa, J. P. , Ramos, C. C. O. , Souza, A. N. , & Papa, L. P. (2013). Optimizing feature selection through binary charged system search. In Proceedings of 15th international conference on computer analysis of im- ages and patterns (pp. 377–384) . Rodrigues, D., Yang, X.-S., Souza, A. N., & Papa, J. P. (2015). Binary flower pollina- tion algorithm and its application to feature selection. In X.-S. Yang (Ed.), Recent advances in swarm intelligence and evolutionary computation . In Studies in Com- putational Intelligence: vol. 585 (pp. 85–100). Springer International Publishing. doi: 10.1007/978- 3- 319- 13826- 8 _ 5 . afont, G. , Salazar, A. , Soriano, A. , & Vergara, L. (2012). Combination of multiple de- tectors for eeg based biometric identification/authentication. In Proceedings of the IEEE international Carnahan conference on security technology (pp. 230–236) . Sanei, S. , & Chambers, J. (2007). EEG signal processing . Chichester, England; Hoboken, NJ: John Wiley & Sons . Schalk, G. , McFarland, D. , Hinterberger, T. , Birbaumer, N. , & Wolpaw, J. (2004). BCI20 0 0: A general-purpose brain-computer interface (BCI) system. IEEE Trans- actions on Biomedical Engineering, 51 (6), 1034–1043 . ubasi, A. (2007). EEG signal classification using wavelet feature extraction and a mixture of expert model. Expert Systems with Applications, 32 (4), 1084–1093 . Tau, G. Z. , & Peterson, B. S. (2009). Normal development of brain circuits. Neuropsy- chopharmacology, 35 (1), 147–168 . ang, Y. , Gao, S. , & Gao, X. (2005). Common spatial pattern method for channel selection in motor imagery based brain-computer interface. In Engineering in medicine and biology 27th annual IEEE conference (pp. 5392–5395) . ilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics Bulletin, 1 (6), 80–83 . ang, X.-S. (2010). Firefly algorithm, stochastic test functions and design optimisa- tion. International Journal Bio-Inspired Computing, 2 (2), 78–84 . ang, X.-S. (2012). Flower pollination algorithm for global optimization. In Proceed- ings of the 11th international conference on unconventional computation and nat- ural computation . In UCNC’12 (pp. 240–249). Berlin, Heidelberg: Springer-Verlag. doi: 10.1007/978- 3- 642- 32894- 7 _ 27 . ang, Y. , Kyrgyzov, O. , Wiart, J. , & Bloch, I. (2013). Subject-specific channel se- lection for classification of motor imagery electroencephalographic data. In Acoustics, speech and signal processing (ICASP), 2013 IEEE international conference (pp. 1159–1169) . http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0001 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0002 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0002 http://dx.doi.org/10.1109/TIFS.2014.2308640 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0004 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0005 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0006 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0007 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0007 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0008 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0009 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0010 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0010 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0011 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0012 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0013 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0014 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0015 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0016 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0016 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0016 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0016 http://dx.doi.org/10.1109/TBME.2014.2317881 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0018 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0019 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0020 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0021 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0022 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0023 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0024 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0025 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0026 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0027 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0028 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0029 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0029 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0029 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0029 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0029 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0030 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0030 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0030 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0030 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0030 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0030 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0031 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0031 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0031 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0031 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0031 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0031 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0031 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0032 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0032 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0032 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0032 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0032 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0033 http://dx.doi.org/10.1007/978-3-319-13826-8_5 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0035 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0035 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0035 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0035 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0035 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0035 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0036 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0036 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0036 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0036 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0037 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0037 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0037 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0037 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0037 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0037 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0037 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0038 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0038 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0039 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0039 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0039 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0039 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0040 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0040 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0040 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0040 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0040 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0041 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0041 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0042 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0042 http://dx.doi.org/10.1007/978-3-642-32894-7_27 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0044 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0044 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0044 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0044 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0044 http://refhub.elsevier.com/S0957-4174(16)30287-1/sbref0044 EEG-based person identification through Binary Flower Pollination Algorithm 1 Introduction 2 The EEG signal 3 Related work 4 Proposed method 4.1 Autoregressive model 4.2 EEG channel selection 4.2.1 Flower pollination algorithm 4.2.2 Binary flower pollination algorithm 4.3 Optimum-path forest classifier 5 Methodology 5.1 Dataset 5.2 Nature-inspired meta-heuristic algorithms 5.3 Experimental setup 6 Experimental results 6.1 Transfer function analisys 6.2 Discussion 7 Conclusions and future work Acknowledgments References