Avatar Navigation in Second Life using Brain Signals

Mehrnaz Kh. Hazrati* Ulrich G. Hofmann

Graduate School for Computing in Neuroelectronic Systems, Department of Neurosurgery,
Medicine and Life Sciences, University Medical Center Freiburg
Institute for Signal Processing, University of Lübeck Freiburg, Germany
Lübeck, Germany ulrich.hofmann@coregen.uni-freiburg.de
hazrati@isip.uni-luebeck.de

Abstract—This paper presents an interdisciplinary project Employing scientific, yet game-based approaches and
conducted in a sophisticated approach to set up a connection technologies constructs also another aspect of new BCI
between the human brain and a virtual world illustration. The generation. Impairment in motor function leads to forseeable
main purpose of this research is to develop an interaction
difficulties in social interactions and a reduction in quality of
life for most patients. Paralyzed people not only suffer from
technique for entertainment and to investigate the virtual reality
aspects for physical rehabilitation. Users could freely explore
Second Life’s virtual world and steer their avatar to the target physical inabilities but also psychological aspects worsening
point using their brain signals. The experiment was carried out over time. Providing a fun and attractive interface might
on three subjects. The control can be accomplished by online provoke the user’s intention in a more effective way.
processing of single-trial EEG signals recorded from EPOC Improving the quality of life for paralyzed people due to spinal
headset, Emotiv, and passing the parameters to the main cord damage, stroke, or peripheral nervous system disabilities
interface. An adaptive filter-based approach was used to reduce are the major clinical motivation for developing a brain-
the eye blink distortions. Time–frequency features with computer interface which may help to restore their mobility
combination of an ensemble LDA classifier led to a maximum through prosthetic devices or software interfaces [11]. Elderly
87.2% of accuracy.
people with stroke [6] and children with sensorimotor
Keywords—Virtual World Environment, Brain Computer disorders may also gain from interactive computer play for
Interface, Second life, Emotiv. rehabilitation purposes [11] [8]. In [5] the effectiveness of the
Wii system was tested for balance rehabilitation in patients
I. INTRODUCTION with acquired brain injury.
A recent research study extends the usage of the well-known
Virtual social worlds (e. g Second Life) are becoming gait rehabilitation robot “Lokomat®” from a boring and
increasingly interesting for people from different generations continuous physical practice sessions to an exciting game
and social groups. Among all these, people with handicaps controller for children [7].
could also benefit broadly from this new technology. In this project a heuristic interaction technique was
Forfeiture of movement control in human is one of the developed that allows the user to freely control an avatar in the
common diseases worldwide with a deeply negative effect on Second Life virtual environment by using mental activities.
the patients’ quality of life and their communication ability [8, The system is trained for some predefined movements. The
9]. Brain-computer interfacing (BCI) promises a direct control can be accomplished by online processing of single-
communication solution from the brain to the external world trial EEG signals which is recorded using the EPOC headset
without employing the peripheral nervous system [1, 3]. The from Emotiv.
use of BCI, however, is still very limited to some serious
clinical applications [2]. In the past few years, emerging II. EXPERIMENTAL SETUP
simple and cheap BCI systems opened up the market and made Experimental runs were carried out with three able-bodied
BCI appealing for the new and large group of people [8] who volunteer subjects (1 female, 2 males). Only S1 had previous
initially started to use BCI for fun and entertainment purposes experience with BCI. In preparation for each run, subjects were
[10]. It expands the yet limited BCI application to more seated 0.5 meters from a computer monitor with arms in front
applicable scenarios in everyday life, even though noninvasive and palms flat and facing down on a desk. The setup consists of
methods will probably have limited use in fine tuned controle hardware and software parts.
[2, 4]. The goal of using BCI in the life of healthy people is to
render it a useful tool in situations where the normal and A. Hardware: Emotiv EPOC headset
conventional interfacing means such mouse, keyboard or The EEG acquisition machine utilized was an Emotiv
joystick cannot be an option, i.e. in space or marine projects EPOC headset. This headset is a low-cost, commercial
[1] or to improve attention, concentration or motivation [7]. To EEG cap with dry electrodes. It is easy to wear and use and is
date prototype BCI applications are available in the domain of equipped with Bluetooth connectivity to the computer. Fig. 1
gaming and entertainment using easy to wear hardware [10]. shows the headset and positions of electrodes covered by it.

978-1-4673-4544-6/13/$31.00 ©2013 IEEE

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on April 29,2021 at 18:50:45 UTC from IEEE Xplore. Restrictions apply.
 One of three images – a left arrow, a right arrow, and a
forward arrow – was shown for five seconds, and the subject
was asked to perform a corresponding brain activity, e. g.
imagination of right hand, left hand of foot movement. After
five seconds, the picture disappeared, and a screen with only
the focal point was shown, indicating that the subject should
relax. The offline run consisted of 120 trials to train three
movements (Left, Right and Forward commands).
Fig. 1. EPOC headset and correponding electrode locations C. Software: Second Life interaction
Since 2007, the Emotiv EPOC is available in the market. Second Life’s virtual world is a well-known web-based
Recently it found great attention from research for BCI based interaction environment [14]. Users can freely explore Second
games [12, 13]. Two reference channels in addition to 14 EEG Life’s virtual world and steer their avatar to the target point
channels are positioned according to the 10-20 system at CMS normally with mouse or arrow keys. A sophisticated new
and DRL and AF3, AF4, F7, F3, F4, F8, FC5, FC6, T7, T8, brain-computer interface based on the functionality of this
P7, P8, O1, O2. Sampling rate is 128Hz. EPOC also includes environment was developed. Second Life has been controlled
gyroscopes to monitor head movement (X and Y axis). using BCIs before, but only to a very rudimentary level [15].
Instead of arrow keystrokes on the keyboard for controlling the
B. Paradigm
avatar, commands were sent directly to the interface. The
Our experimental paradigm consisted of one offline session interface is developed in C++ using Emotiv's SDK Developer
for training the classifier followed by one online classifier Edition. It passes the arguments to the main interface using the
evaluation / virtual world control run depicted in Fig. 2. keybd_event() Win32 API function which is recognized as a
During each offline trial an image of a black circle was keystroke. It executes the corresponding command in the
continuously displayed in the middle of the screen for use as a virtual world. Thus, it lets users freely explore Second Life’s
visual fixation point to minimize eye movement artifacts. Each virtual world and control their avatar within it. During the
trial lasted ten seconds, the first five were devoted to the online mode, the system allows the user to decide with some
imagination of a particular limb movement showing the arrows degree of freedom to choose the direction of movement and
in Fig. 2 and the last five seconds were dedicated to the steer their avatar to the target point.
relaxation mode.
D. Evaluation of the data
In order to evaluate the recorded data, some further offline
analyses have been accomplished. The grand average ERPs for
this dataset were calculated. Fig. 3 presents grand average for
AF3, AF4, F4, F3, FC5, FC6, P7 and P8 for trials associated to
the right, left, forward arrow and stop (idle state). Units on x-
axis are ms which represent the first 1s of the entire trial
starting from the point when the arrow appeared on the screen
at the beginning of each trial. The entire dataset includes right,
left, forward and stop commands.
The first deflection is the negative peak of the evoked
potential, consisting of a -4.1 µV and -4.5 µV peak at about
B 220 ms, maximum at P7 and P8. The maximum amplitude
decreases from parietal to frontal sites. All over recorded
channels, especially at F3, the grand average shows another
positive peak after 400 ms, most probably a P300 component
according to its timing and topography. This can be seen for
all arrow-based commands and decays for the stop command,
when no sign appeared on the screen. This is followed by a
-5.8 µV negative peak 500 ms after the start of the trial,
maximum at F3 and P7, especially for the right arrow
command. It happens 50 ms earlier at P7. Fig. 4 illustrates the
A
average ERP distribution 300ms after the onset of experiments
and each second after this time point at 1300 ms, 2300 ms,
3300 ms and 4300 ms. The maximum value can be seen in the
Fig. 2. Second Life snapshot. The user is trained to command an avatar by middle of trials, this may be the result of observing some
thinking about the direction commands: Left, Right and Forward. During the
movement (biofeedback) which happened usually in the
test section user should navigate the avatar from point A to point B.
beginning of the trial for the majority of subjects

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on April 29,2021 at 18:50:45 UTC from IEEE Xplore. Restrictions apply.
 III. MATERIALS AND METHODS
Grand average-AF3
Grand av erage-AF4
10
15
forward
8 right
left
forward
right A. Preprocessing
6 10 left
stop stop
Preprocessing consisted of filtering and artifact removal.
Amplitude (uV)

Amplitude (uV)
2 5
The signals were filtered using an eleventh-order zero-phase
0

-2
0 forward and reverse Chebyshev lowpass filter with cutoff
-4
-5
frequency of 100 Hz. A notch filter at 50Hz was applied to
-6 remove power line noise. In each trial the mean values over
-8
100 200 300 400 500 600 700 800 900
-10
100 200 300 400 500 600 700 800 900 time of the imagination and relaxation signals were calculated
Time (ms) Time (ms)
Grand average-F3 Grand av erage-F4 and subtracted from each respective signal. In order to
8 15
forward forward implement an artifact removal system online, we used an
6 right right
left 10 left adaptive filter approach in the time domain. The data of the
4 stop stop
offline experiment was used to initialize the classifier and to
Amplitude (uV)
Amplitude (uV)

5
2
calculate the normalization and artifact reduction parameters.
0
0

-2
-5
B. Feature extraction and selection
-4
As a first step, EEG recordings from each channel are
-6 -10
100 200 300 400 500 600 700 800 900
Time (ms)
100 200 300 400 500 600 700 800 900
Time (ms)
decomposed into a sum of atoms, each being the product of
8
Grand average-FC6
4
Grand av erage-FC5 spatial (topographic) signature and waveforms having
forward
right 3
forward
right
determined time–frequency localization. Matching pursuit was
6
left
stop
2
left
stop
exploited to find the best matching projections of
4
multidimensional data onto an over-complete dictionary of
Amplitude (uV)

Amplitude (uV)

1
2 0
Gabor functions. The five first weighting coefficients extracted
-1
0
-2
from each channel were used as features in BCI.
-2
-3
-4
-4 1) Matching pursuit
-6 -5
100 200 300 400 500 600 700 800 900
Time (ms)
100 200 300 400 500 600 700 800 900
Time (ms)
Matching pursuit (MP) was initially introduced by Mallat
5
Grand average-P7
6
Grand average-P8 and Zhang and to date there are several extensions to the
forward
right
forward
right
original formulation [16] [20]. MP provides an intuitive and
4
left
stop
left
stop
rather simple algorithm which refers to a greedy
Amplitude (uV)

decomposition of signals into basic waveforms from a very

Amplitude (uV)

0 0 large and redundant dictionary of functions. The robustness of

-2
the algorithm depends on the complexity of the combination
-4
of dictionary’s functions [16] [18]. It forms an adaptive time-
frequency representation of the signal and is based on an over-
-5 -6
100 200 300 400 500 600 700 800 900
Time (ms)
100 200 300 400 500 600 700 800 900
Time (ms)
complete dictionary [20]. Each iteration of the algorithm
searches for the most informative projection, yielding an
Fig. 3. Grand average- ERP calculated for the Dataset. Averaged over all
trials associated to the right arrow recording when at the beginning of trial a estimate which is basically free from arbitrary settings.
appears. The entire dataset includes right, left, forward and stop commands. Matching pursuit is a type of numerical technique which tries
to find the best matching projections of multidimensional data
onto an over-complete dictionary D. The basic idea is to
300 ms 1300 ms 2300 ms
represent a signal f(t) from Hilbert space H, as a weighted sum
of functions (called atoms) taken from D. A common
choice for functions are Gabor functions, which is a
sinusoidal modulated by a Gaussian.

3300 ms 4300 ms
3.1

Then the signal f(t) can be decomposed as:

1.6

-1.6

-3.1
where n indexes the atoms that have been chosen, and is
at1
the weighting factor (an amplitude) for each atom. Given a
Fig. 4. The average ERP distribution 300ms after the onset of experiments fixed dictionary, matching pursuit will first find the atom that
and every second after that time point.
has the biggest inner product with the signal, and then subtract

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on April 29,2021 at 18:50:45 UTC from IEEE Xplore. Restrictions apply.
the contribution due to that atom; this process is repeated until Basis
the signal is satisfactorily decomposed [16]. 1

0
TABLE I. ALGORITHM MATCHING PURSUIT
-1
0 50 100 150 200 250 300 350 400 450
1.Initialization:
10 signal
approximation

2. Find with maximum inner product 0

3. Let -10
0 200 400 600 800 1000 1200 1400

increasing frequency ->

4. Let
weights (50 non-zero weights)
5.Let
5
4. If not converged ( ), go back to 2
10

We would like to use a dictionary D= { } n=1…N that 15

200 400 600 800 1000 1200
would properly reveal the intrinsic properties of an unknown
signal, or, almost equivalently, would give low entropy of
the { }. Unfortunately, there is no universal recipe for such a Fig. 5. Top plot: Atoms used in the matching pursuit method. It is generated
prior choice. The algorithm is summarized in the Table I. by changing the Gabor function parameters in order to generate a complete
dictionary. Middle figure: The recorded and the approximation signal
f(t)and dictionary D are inputs, and the output is the list calculated by Matching pursuit decomposition for one EEG channel (F4-5s)
of coefficients ( , ). recorded from S2. Bottom figure: feature set consists of values of 50 non-zero
Fig. 5 shows a set of features extracted using this method. The weights of used Gabor atoms.
top plot presents atoms used in the matching pursuit method. It
is generated by changing the Gabor function parameters in In our BCI we used the mRMR method in order to decrease
order to generate a complete dictionary. Middle figure: The the feature space to 10 features as the input of each classifier.
recorded and the approximation signal calculated by MP The results confirm that mRMR leads to higher classification
decomposition for one EEG channel (F4-5s) recorded from S2. accuracy and increase the speed of online processing as well.
Bottom figure: feature set consists of values of 50 non-zero
weights of used Gabor atoms. In order to decrease the C. Linear Discriminant Analysis (LDA) Classification
calculation time during the online experiment we applied a
From the mathematical point of view the aim of each
feature selection method to increase both the speed and the
classification algorithm is to define a function for the decision
accuracy of the system.
boundary between two or more classes. In a multidimensional
2) Maximum relevance minimum redundancy (mRMR) feature space the decision boundary is indeed a separating
Selecting a subset of superior features using the maximal hyper-plane [18]. When these decision boundaries are linear;
statistical dependency criterion is a well-known feature methods for classification are linear. Linear discriminant
selection method in BCI [22]. Max-Dependency criterion is analysis (LDA) is a classic classifier, as its name suggests,
based on mutual information and it is mathematically difficult with a linear decision surface [18]. LDA classification is
to calculate this value in a real-world application. Here we parameter free and has a closed form solution that can be
applied an alternative approach for selecting features based on easily computed. Fig. 6 demonstrates a typical two-dimension
a maximal relevance minimum redundancy criterion, the boundary solution for the depicted dataset. The LDA approach
mRMR algorithm. The method uses the mutual information to is based on finding the best linear hyper-plane for separating
select (M=10) features that best fulfill the minimal redundancy the data of the different classes [17]. In our four-class
and maximal relevance criterion at the same time. It calculates problem the three lines form the boundaries.
the mean value of all mutual information values between Assume is the between-class variance matrix and is the
individual feature and class c over the offline dataset. It is within-class variance matrix calculated as:
likely that features selected according to Max-Relevance could
have rich redundancy, i.e., the dependency among these
features could be large. When two features highly depend on
each other, the respective class- discriminative power would
not change much if one of them were removed. Therefore, the
following minimal redundancy (Min- Redundancy) condition The optimal solution is based on finding the optimum weight
can be added to select mutually exclusive features [21]. A
vector so that the mean of two classes has maximum
detailed description of the mRMR method can be found in
distance from each other and the variance between each class
[21].

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on April 29,2021 at 18:50:45 UTC from IEEE Xplore. Restrictions apply.
is minimized. It is then calculated as the eigenvector FP and FN are also called error type I and II respectively [17].
corresponding to the maximum eigenvalue of [19]. FP is incorrect rejection of a positive result and FN is the
failure to reject a false result [17].

A receiver operating characteristic (ROC) curve is a standard

After calculating the training feature vector, each test feature evaluation of the classifier that plots sensitivity versus
vector can be classified employing a simple decision rule such specificity. Sensitivity is a measure of evaluation which
as a sign function. Training was repeated for 10 runs each measures the percentage of TP divided by the sum of
containing 12 trials for three movement imagination misclassification both in the form of TP and FN. Specificity is
a measure of evaluation which measures the percentage of TN
divided by the sum of misclassification both in the form of TN
and FP. Fig. 8 demonstrates the result of ROC analysis for the
stop and movement commands during the experiment for all
subjects. ROC curve shows an optimal balance of false
positive and false negative at an area under curve (AUC) of
more than 0.887. By changing the classifier parameters the
sensitivity and specificity change. The optimum classifier is
selected by looking at the right upper corner of the ROC
curve. Table II shows the average error rate for the first
session in three subjects. False Negative (FN) commands are
Fig. 6. Fisher’s LDA is a simple linear classifier based on finding the hyper-
plane to separate the data representing the different classes. minimized by using an ensemble classifier, which is designed
based on the selected features from all channels.
D. Ensemble classification ROC curve (AUC=0.8807)
1
In our work we exploit the concept of combining several
classifiers which is usually called ensemble classification. 0.9
Three LDA classifiers for each channel were trained (Left, 0.8
Right, Forward vs. Stop). The final decision was made based
on the output of all classifiers. 0.7

0.6
Sensitivity

0.5
Feature Classifier 0.4
ext. Ch.1 .1
0.3

0.2
Feature Classifier
Decision
ext. Ch.2 .2 0.1
making
0
0 0.2 0.4 0.6 0.8 1
Specificity
Feature Classifier
.n Fig. 8. ROC curve for the relaxed and movement accuracy during the
ext. Ch.n experiment for all subjects. By changing the classifier parameters the
sensitivity and specificity change. The optimum classifier is selected by
Fig. 7. The idea of ensemble classification for different channel and decision looking at the right upper corner of the ROC curve
making based voting

TABLE II. THE AVERAGE OF ERROR RATE AND ACCURACY OF THE

IV. RESULTS AND DISCISSION CLASSIFICATION IN TRAINING AND TEST SESSION FOR THREE SUBJECTS
The efficiency of a BCI, in terms of recognition rate, Evaluation (% in average)
depends mostly on the choice of appropriate features and Method
(LDA) Training Test
TP TN FN FP
classifiers. Temporal-frequency features with combination of session session
an ensemble LDA classifier was used for decision making. S1 43 37 16 7 87 76
The command was extracted from 14 channels EEG signal
S2 38 33 17 12 79 65
each second. For a better evaluation of the entire system we
calculated some further parameters in our dataset. These S3 40 34 16 10 81 77
concepts are defined based on True Negative (TN), True
Positive (TP), False Negative (FN) and False Positive (FP).
TN is when the test makes a negative result (stop command) Fig. 9 shows the classification accuracy of the stop and move
and the subject actually intended it. Similarly, TP is when both (Right, Left or Forward) commands in several runs of the
test and the subject have positive results (movement
command).

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on April 29,2021 at 18:50:45 UTC from IEEE Xplore. Restrictions apply.
online experiment for S1, S2 and S3. S1 with previous experiment. Each feature extracts an important but selective
experience in BCI showed the most stable results. aspect of mental tasks from the dataset. In our experiments the
best classification accuracy was obtained when a subset of 10
100 features were combined with ensemble LDA classifer.
90

80 Our experiment is one of the few times the latest BCI

Classification accuracy (%)

technology has been combined with web-based virtual reality

70
technology and online gaming. Our proposed method uses
60 electroencephalography (EEG) caps to pick up brain signals,
50 and translate them into commands that are relayed to control a
virtual avatar in Second Life. It can be combined with a real
40
world environment too: turning on and off the computer and
30 opening other web pages or answering the phone are potential
20 improvements.
Ralax
10 Event ACKNOWLEDGMENT
Total
0 This work was supported by the Graduate School for
1 2 3 4 5 6 7 8 9 10
Run : AH, Session : 3
MLP Classifier, Subject Computing in Medicine and Life Sciences funded by
100 Germany's Excellence Initiative [DFG GSC 235/1].
90
REFERENCES
80
Classification accuracy (%)

70 [1] Thomas D. Coates, Jr., Neural Interfacing: Forging the Human-Machine

60 Connection, Morgan & Claypool Publishers series, 2008
[2] Schalk, G. et al., Decoding Two-Dimensional Movement Trajectories
50 Using Electrocortico-graphic Signals in Humans, Journal of Neural
40 Eng, 4: 264-275, 2007.
[3] Nijholt A, Tan D , Playing with your brain: Brain-computer interfaces
30 and games. In: Proceedings of the International Conference on Advances
20 in Computer Entertainment Technology. ACM, New York, NY, USA,
Ralax pp 305–306, 2007.
10 Event [4] Wolpaw, J. R. & McFarland, D. J. ,”Control of a two-dimensional
Total movement signal by a noninvasive brain-computer interface in humans”,
0
1 2 3 4 5 6 7 8 9 10 Proc. Natl Acad. Sci. USA 101, pp. 17849–17854, 2004
Run : AZ, Session : 17
RPNN Classifier, Subject [5] Sandlund M, McDonough S, Häger-Ross C., Interactive computer play
100 in rehabilitation of children with sensorimotor disorders: a systematic
90
review. Dev Med Child Neurol. vol. 51, no. 3, pp:173-9. Jan 2009.
[6] Gil-Gomez et al, “Effectiveness of Wii balance board system for balance
80 rehabilitation in patients with acquired brain injury”, Journal of
Classification accuracy (%)

70
NeuroEngineering and Rehabilitation, vol. 8 no. 30, 2011.
[7] Schuler T., Brutsch K., Muller R., Hvan Hedel U. J., Meyer-Heim A.,
60 Virtual realities as motivational tools for robotic assisted gait training in
50
children: A surface electromyography study, Neuro Rehabilitation, vol.
28, no. 4, pp: 401-411, 2011
40 [8] Brooks Jr., F. P. , “What's Real About Virtual Reality?”, IEEE
30
Computer Graphics And Applications, vol. 19, no. 6, 16, 1999.
[9] Kempe L., “Serious Games für die Gesundheit: Spielen für eine ernste
20 Sache“, Dtsch Arztebl, vol. 107, no. 38, 2010
Ralax
10 Event [10] Roman Krepkiy, Brain-Computer interface: Design and implementation
Total of an online BCI system for the control in gaming application an virtual
0 limb, VDM Verlag Dr Muller, 2008
1 2 3 4 5 6 7 8 9 10
Run [11] Keshner Emily A, Virtual reality and physical rehabilitation: a new toy
or a new research and rehabilitation tool?, Journal of Neuroengineering
Fig. 9. Performance of S1, S2 and S3 in several runs of the online and rehabilitation, vol. 1, no. 8, 2004.
experiment. The red line corresponds to the stop command and the green line
shows the classification accuracy in average for move commands (Right, Left [12] Gonzalez-Sanchez, J., Chavez-Echeagaray, M. E., Atkinson, R., and
Burleson, W., An Agent-Based Software Architecture for a Multimodal
or Forward). The Black line demonstrates the average performance for all four
Emotion Recognition Framework. In Proceedings of 9th Working
classes.
IEEE/IFIP Conference on Software Architecture (WICSA’11), 2011.
[13] Stefan Debener et al, How about taking a low-cost, small, and wireless
On average there was a decrease in performance when EEG for a walk?”, Psychophysiology, vol. 49, no. 11, pp: 1617-21,
subjects encountered the 3D interface after the training PMID 23013047, 2012
session. The reason could be that we have selected the best [14] Au, Wagner James. The Making of Second Life, pg. 19. New York:
Collins. ISBN978-0-06-135320-8.
features in training data and applied the same subset during the
online

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on April 29,2021 at 18:50:45 UTC from IEEE Xplore. Restrictions apply.
[15] Alexander Y. Kaplan et al., Adapting the P300-based brain-computer [20] Mallat S. G. and Zhang, Z. Matching Pursuits with Time-Frequency
interface for gaming: a review, IEEE Transactions on Computational Dictionaries. IEEE Transactions on Signal Processing. December 1993,
Intelligence and AI in Games (Special Issue on Brain/Neuronal- pp. 3397–3415.
Computer Games Interfaces and Interaction), 2013, in press. [21] Peng H, Long F., and Din Ch. Feature selection based on mutual
[16] Durka P. J., Dobieslaw, I. and and Blinowska, K. J. Stochastic time– information: criteria of max-dependency, max-relevance, and min-
frequency dictionaries for matching pursuit. IEEE Trans. Signal redundancy. IEEE Transactions on Pattern Analysis and Machine
Process., vol. 49, no. 3, March 2001 Intelligence, vol. 27, no. 8, pp. 1226-1238, 2005.
[17] Hastie T., Tibshirani R., Friedman J. The Elements of Statistical [22] Koprinska, Irena Comparison of Feature Selection Methods for
Learning. 2008 Classification of Brain-Computer Interface Data. Workshop on Advances
[18] Saeid Sanei and Chambers J. A. EEG signal processing. s.l. : John and Issues in Biomedical Data Mining, pp. 41-50, 2009.
Wiley & Sons Ltd., 2007
[19] Fukunaga K. Statistical Pattern Recognition. New York: : Academic,
2nd edition, 1990
.

Authorized licensed use limited to: Escuela Superior de Ingeneria Mecanica. Downloaded on April 29,2021 at 18:50:45 UTC from IEEE Xplore. Restrictions apply.