Real Time Drunkness Analysis Through Games Using Artificial Neural Networks

Abstract

In this paper, we describe a blood alcohol content estimation prototype based on a comportment analysis performed by artificial neural networks. We asked to subjects that had drunk alcohol to play a video-game after having measured their blood alcohol content with a breathalyser. A racing game was modified so that it could provide various data related to the use of the controls by the player. Using the collected data, we trained our neural network in order to be able to determine whether or not the subject had exceeded a blood alcohol content threshold. We also succeeded in estimating this blood alcohol content with a mean error of 0.1g/l. We could perform those estimations independently of the track played among the two ones used. It was also performed in "real time", e.g., using only the data collected within the last 10 seconds of playing.

Full text

Real Time Drunkness Analysis Through Games Using
Artificial Neural Networks
Audrey Robinel, Didier Puzenat
To cite this version:
Audrey Robinel, Didier Puzenat. Real Time Drunkness Analysis Through Games Using Artifi-
cial Neural Networks. International Conference on Advances in Computer-Human Interactions,
Feb 2011, Gosier, Guadeloupe. pp.206-211. <hal-00849971>
HAL Id: hal-00849971
https://hal.archives-ouvertes.fr/hal-00849971
Submitted on 2 Aug 2013
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entific research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destin´ee au d´epˆot et `a la diffusion de documents
scientifiques de niveau recherche, publi´es ou non,
´emanant des ´etablissements d’enseignement et de
recherche fran¸cais ou ´etrangers, des laboratoires
publics ou priv´es.

Real Time Drunkness Analysis Through Games Using Artificial Neural Networks
Audrey ROB IN EL
LAMIA, Universit´
e Antilles-Guyane
Pointe-`
a-Pitre, France
audrey.robinel@univ-ag.fr
Didier PU ZE NAT
LAMIA, Universit´
e Antilles-Guyane
Pointe-`
a-Pitre, France
didier.puzenat@univ-ag.fr
Abstract—In this paper, we describe a blood alcohol content
estimation prototype based on a comportment analysis per-
formed by artificial neural networks. We asked to subjects that
had drunk alcohol to play a video-game after having measured
their blood alcohol content with a breathalyser. A racing game
was modified so that it could provide various data related to
the use of the controls by the player. Using the collected data,
we trained our neural network in order to be able to determine
whether or not the subject had exceeded a blood alcohol content
threshold. We also succeeded in estimating this blood alcohol
content with a mean error of 0.1g/l. We could perform those
estimations independently of the track played among the two
ones used. It was also performed in “real time”, e.g., using only
the data collected within the last 10 seconds of playing.
Keywords-User interfaces; Games; Neural network applica-
tions; Cognitive sciences; Psychology; Human factors.
I. INT ROD UC TI ON
Driving under influence is not only forbidden, but also
dangerous, being a major cause of accidents. Devices for
measuring the blood alcohol concentration of a driver have
thus been developed. The current and most common ap-
proach for a driver is to use a “breathalyser” [3]. Such a
device measures the amount of alcohol contained in the
exhausted air blown by the subject into the device by using
a chemical reaction. The blood alcohol content is computed
from this value. Law enforcement class devices are very
reliable but expensive. Most of the low cost devices sold
to regular citizen are unreliable and provide erroneous mea-
sures, and require regular calibration or parts replacement
(we had a hard time finding a good device, which came at a
higher price). Furthermore, if ever a good device is used, the
driver must not forget to test himself before driving, which
is very likely to happen for a drunk person.
But what if the car could estimate whether or not the
driver is in condition for driving? Alcohol affects the driver
and thus creates dangerous behavior on the road. This means
that if we monitor the “low level” characteristics of the car
controls used, we should be able to correlate the subject’s
blood alcohol content to his driving ability [9]. Our final
goal is to create a natural interface embedded in the vehicle
that would use the data collected by the car’s computer
to analyze the driver’s behavior and actions to provide a
diagnostic to the driver and warn him before he becomes
really dangerous. For now, we present a prototype in order
to demonstrate the feasibility of the concept. We will not
use a real car but a video game, and gather data from the
use of the game controls. In order to fulfill the objectives
(detailed in the Section II), we had to instrument a game
(Section III-A). We chose SuperTuxKart, a racing game for
the reasons presented in the Section III-C. Following the
measuring protocol described in Section III-E, we collected
low level data (Section III-D) of 120 game runs. Using
all those examples, and a K-fold cross-validation derived
algorithm (Section IV-A), we measured the ability of the
network to detect whether or not a subject’s blood alcohol
content exceed a threshold in Section V. We thereafter
used the system to estimate the value of the blood alcohol
content, in Section VI. We then showed how the system
could perform this task independently of the track played,
in Section VII. Subsequently, after defining what we call a
“real time estimation” in Section VIII, we demonstrate the
ability of our prototype to perform such an estimation in
Section VIII. To conclude, we present in Section IX a global
interpretation of the results and the perspectives opened.
II. OB JE CT IV ES
Our main goal is to demonstrate that using the data
collected from an instrumented game played by a subject, a
trained artificial neural network is able to determine whether
or not a subject’s blood alcohol content exceeds a fixed
value, and even estimate the blood alcohol content value.
Our secondary goal is to determine if the neural network is
able to provide those estimations independently of the track
or if it must be used on predefined tracks. Our last goal for
the prototype is to perform in real time (we have a specific
definition, presented in Section VIII), at any moment of the
game run.
III. OUR S YS TE M
A. “Instrumenting” a game
We instrumented a video game that provided many low
level measurements. By “Instrumenting” we mean taking
measurements of player actions on input devices such as
a joystick or a steering wheel. This devices response being
progressive rather than binary, we can collect the state of
the device and obtain a large set of “low level” data about
the intensity of the subject’s actions on the controls. We

gather these data continuously in time, which enables us to
compute the evolution of some parameters over time.
B. Artificial neural networks
Playing a video game implies mobilizing many skills.
Depending on the game, it may be more or less necessary
to think, to act quickly, to act with precision, etc. Thus,
instrumenting a game should give valuable information on
the player. However, the point is to know what to measure
and how to go from the measurements to a conclusion.
The prototype presented in this paper uses artificial neural
networks to select and combine relevant measures to assess
the player, and is based on previous works [9]. Once
trained on a representative population, a dedicated neural
network will be able to evaluate a new player. Our approach
eliminates the need to explicit the significance of measures
for a given goal, therefore it becomes possible to make “low
level measurements”, e.g., the frequency of corrections on
the steering wheel during a racing car simulation.
Among all artificial neural network models, for all our
present developments, we have chosen a classical Multi-
Layer Perceptron (MLP) with a back-propagation learning
algorithm [2]. The neural network has been implemented
with the free open source Fast Artificial Neural Network
Library (FANN) [8]. FANN is a C library facilitating the
use of the developed neural network within the instrumented
video game (or interface, embedded computer, etc.), e.g., to
perform a real-time evaluation.
C. The chosen game
We chose a racing game because of the dynamic aspect
of such games. Since races are played on tracks, it implies a
predefined path. The subject has then a restricted freedom,
which turns the runs (played games) into reproducible and
comparable tasks, enabling us to create an uniform exam-
ples set. Furthermore, it can be controlled with a steering
wheel [7], which is an intuitive device for controlling a car
as presented in Figure 1.
We selected SuperTuxKart [5] because it is extremely easy
to handle. Anyone who could drive a car, even if unfamiliar
with computers can perform the tests. It has very simple
tracks, on which it is impossible to loose one’s way (every
subject were able to finish the races). Tanks to the open
source license (GNU GPL), we could edit the code to include
our data collecting library.
D. The collected data
For each full lap of the circuit, a vector of 5 components
is preserved and will feed the artificial neural network.
Those components are (i) the average number of steps per
rotation (the steering wheel is analogical), (ii) the number
of accidents (collisions and falls), (iii) the total number
of actions of the steering wheel, and (iv) the number of
changes in direction (e.g., the driver was turning left and is
Figure 1. The prototype during a run
now turning right). We define a “run” as a session played
by a subject on a full track lap. When a subject plays a
run, the system generates XML files containing all the race
measurements. Each run then provides us an example (in the
real time context, each run will provides us with multiple
examples, more on that in Section VIII) that can be used for
our problematics. Each subject does multiple runs, and the
combination of all runs is our full examples set. For each
run, we also measured the subject’s blood alcohol content
with a breathalyser.
E. The measuring and data collection protocol
In order to obtain a coherent example set, we used
the same protocol for each subject. We had selected two
tracks (“skyline” and “snowmountain”, see Figure 2) for our
data collection campaign, and every subject played on both
tracks. Each subject played being sober on skyline, then
on snowmountain, so that he could get used to the game
controls. We then repeated it, still with a 0 g/l blood alcohol
content, which provided us sober playing examples. Then the
subject drank some alcohol, and we waited 20 minutes. After
this delay, the subject’s blood alcohol content were measured
with a breathalyser (model AL7000 from Alcopass [1]). Two
other games were then played (same tracks). The operation
were repeated 15 to 20 minutes later, and for some subjects
another time 15 to 20 minutes later again. This provided
us with 8 to 10 examples per subject, two with a blood
alcohol content of 0 g/l, and the other ones with various
values greater than 0 g/l and lesser than 1 g/l.

At last, we obtained a 120 examples set containing game
statistics distributed between 14 different subjects (in order
to avoid subject-specific results) with various blood alcohol
content, between 0 and 1 g/l. We ensured that, after the
experiment, no subject drove a real car while having a blood
alcohol content superior to 0 g/l.
Figure 2. The skyline track on the left, snowmountain on the right.
IV. EXP ER IM EN TAL P ROTO CO L
Before presenting our results, we will explain the method-
ology used to obtain the numbers that will be showed later.
Neural networks are trained on a learning set and then used
in generalization, on examples that are not in the learning
set. Generalization can be done either in production, or in
order to test the network as presented in Section IV-A. We
used a method derived from K-fold cross validation [6]
to evaluate the generalization performance of the network.
This provided the success rate and the average error of the
network for each problematic.
A. The K-fold cross-validation derived algorithm
Our algorithm was the following : at start, we take an
example base (in this example, it has 100 examples). We
initialize the algorithm by setting a loop index i, variable s
(successes) and variable f(failures) to 0. We then split
our example base in two: examples 0, 1 and 2, constitute
the generalization set and examples from 3 to 99 form
the learning set. We train the network on the learning
set, and test it on the generalization set. We evaluate the
network’s answer and increment sor fin case of success
or failure. We then increment i, and start again the same
with generalization set containing examples 1, 2 and 3, and
learning set containing examples 4 to 99 and example 0.
We keep doing that until ireaches 99 (in that case, the
generalization set is constitute of examples 99, 0 and 1,
while learning set is the rest). Thus, on each pass, we trained
the network on a subset of examples, and tested it on another
disjoint subset. Therefore, the network was never tested in
generalization on any example used in the learning set. We
obtain in the end the total number of successes and failures.
We can then compute the success rate of the neural network
as accurately as possible.
B. Determination of right and wrong network answers
We will now explain how we decide when the neural net-
work response is right or wrong. When the network returns
(i) skyline (ii) snow (iii) both tracks
successes 129 121 233
failures 3 5 25
success rate (%) 97.73 96.03 90.31
mean error (g/l) 0.0587 0.0972 0.1873
Table I
THR ESH OL D EXC EE DIN G DE TER MI NATIO N
a value for an example, we can compare it to the real value,
e.g., the estimated blood alcohol content and the measured
value. We define ǫas the distance between the expected
value vand the neural network output o:ǫ=kv−ok.
We define the total cumulated error as the sum of all ǫ
obtained for each example. We then obtain the average error
by dividing this sum by the total number of examples tested.
We will set a maximal tolerated value (ǫmax) for ǫ: if ǫ
exceed ǫmax, the network answer will be considered wrong
(a failure). It will be right (a success) otherwise. A value of
ǫmax is defined for each experiment.
From now on, when we write that the success rate in
generalization is x% for a value of epsilon, it means that for
x% of the tested examples, the error of the network were less
than ǫmax. We always give the success rate in generalization,
and never the learning rate.
V. DETERMINATIO N OF T HR ES HO LD E XC EEDING
We first tried to determine if the neural network is able
to detect when a subject’s blood alcohol content exceeds
a fixed threshold of 0.4 g/l. When the subject’s measured
blood alcohol content was over 0.4 g/l, the network was
expected to return a value of 1 on the output, and a value
of 0 otherwise.
We present in Table I the success rate and mean error of
the network when using an example set of runs played on
(i) skyline, (ii) snowmountain and (iii) both tracks.
This first experiment is the easiest for the network, and
offers the best success rates, over 95% of good determina-
tions. The results are similar for skyline and snowmountain,
while they slightly decrease when we use both tracks in the
same base (it will be studied further in section VII).
VI. ES TI MATI ON O F BL OO D AL CO HO L CO NTENT
This experiment was conducted in order to see if the
neural network could give an estimation of the subject’s
blood alcohol content. The network is expected to return (as
output value) the subject’s measured blood alcohol content,
between 0 g/l and 1 g/l.
We note in Tables II and III lower success rates than in
the previous experiments (especially for lower values of ǫ).
However, this was expected. As we try to return an accurate
value, we define a lower tolerated error (ǫmax). This indeed
decreases the success rate.

ǫsuccesses failures success rate
0.10 85 47 64.39%
0.12 102 30 77.27%
0.14 108 24 81.82%
0.15 112 20 84.85%
0.16 114 16 87.88%
0.18 119 13 90.15%
0.20 122 10 92.42%
Table II
BLO OD AL CO HOL C ON TEN T ES TIM ATIO N ON S KYL IN E,IN F UN CTI ON
OF ǫ. THE AVER AGE E RRO R IS 0.083860 G/L.
ǫsuccesses failures success rate
0.10 70 53 56.91%
0.12 71 52 57.72%
0.14 78 45 63.41%
0.15 81 42 65.85%
0.16 84 39 68.29%
0.18 94 29 76.42%
0.20 98 25 79.67%
Table III
BLO OD AL CO HOL C ON TEN T ES TIM ATIO N ON S NOW MOU NTAI N,I N
FU NCT IO N OF ǫ. THE AVER AGE E RRO R IS 0.1158 G/L.
VII. TRAC K IN DE PENDENT ESTIMATIO NS
In order to demonstrate that the system is able to estimate
the blood alcohol content of the subject independently of the
race track played, we conducted two other experiments. The
first one, in Section VII-A, focuses on a network trained on a
base containing examples from both tracks. The second one,
presented in Section VII-B, shows how the network behaves
when trained on a learning base containing examples from
one track and is tested on a generalization base made of
examples from another track.
A. Testing the network with examples from both circuits
We will start with the mixed tracks data sets. The example
set includes data gathered from game runs played on the
two tracks, skyline and snowmountain. Using K-fold cross
validation, we measured the ability of the network do
estimate the blood alcohol content of the subjects after being
trained on such a base.
The results presented in the Table IV does not show a
significant changes in neither the success rate nor the mean
error of the network. This tends to indicate that the prototype
is able to estimate the blood alcohol content without the
variable value
successes 77
failures 22
success rate (%) 77.78
mean error (g/l) 0.0587
Table IV
RE SULT S OF TH E MI XED E XA MPL E SE T FOR ǫ=0.2 0
values of epsilon A : sky →snow B : snow →sky
Success rate for ǫ=0.15 26.47% 40.62%
Success rate for ǫ=0.20 82.35% 62.62%
Success rate for ǫ=0.25 88.24% 71.88%
Average error (g/l) 0.185407 0.226358
Table V
TRAC K INDE PE ND ENC E TE ST RE SU LTS.
need of a specific portion of road, implying that we could
construct examples set from various road sections or tracks.
B. Training the network with examples from one track, and
testing it with examples from another track
After this, we trained the network on the examples from
the first track, and then checked it’s generalization results on
the examples of the second track. This configuration ensures
that the second set of examples (games played on the second
track) is unknown to the network.
This time, we created two examples sets instead of one: (i)
the “sky set” contains all the game runs that were played on
the “skyline” track; (ii) the “snow set” contains all the game
runs that were played on the track “snowmountain”. We will
try two configurations: in the first one (A) we will use the
sky set as training set and snow set as generalization set. In
the second one (B), the the snow set as training and sky set
as the generalization set. The results obtained for different
values of ǫare presented in Table V. On contrary of other
experiments, we did not use the k-validation algorithm, since
we had two distinct sets.
If we consider low values of ǫ, the success rate is quite
low. We had to increase ǫin order to maintain a success
rate comparable to the previous ones. The obtained ǫreaches
really important max values, with in the worst case 0.25 g/l.
But despite the fact that the mean error also increased,
the network could still perform valuable estimations. This
tends to indicate the ability of the prototype to be used
to estimate the blood alcohol content independently of the
track. Furthermore, this shows how our neural network is
able to cope with data gathered on an unknown portion of
track. This avoids the need to either create an exhaustive
learning set containing all possible situations or to use the
system in the same environment that were used for training.
Of course it is far from being perfect as, in these conditions,
the prototype is less accurate with much lower success rates
in some cases (experiment B, for example) and significantly
higher average errors (reaching 0.22 g/l for experiment B).
But the prototype tends to demonstrate that it is able to
generalize from one track to another.
Of course, we could have more meaningful results with
more tracks, but the game did not provided enough tracks
that met our requirements, detailed in section III-C.

VIII. RE AL T IM E ES TI MATI ON S
We will here demonstrate the ability of the system to
perform “real time” estimations of the measured blood
alcohol content.
A. Definition of “real time estimation”
We define as “real time” an estimation done using the
data gathered within the last elapsed seconds. We define n
as the length in seconds of the interval of play used. An
nseconds real time estimation means that the estimation is
performed using the data of nseconds of race instead of
using the full length of the race. The interval of play used
for this determination will be called a “window” and nis
defined as the size of the window.
In order to obtain comparable data, we normalize the
variables by dividing their values by n. By this mean we
also ensure that the trained network can estimate the blood
alcohol content for various values of n, e.g., it can be used
indifferently to provide a real time estimation based on 10
or 20 seconds.
We kept using the examples from both tracks for our real
time experiments.
B. Protocol
We checked here if the network is able to perform a real
time determination based on 10, 15, 20, 30, 40 and 50 s
windows. In order to study the impact of non the results,
we kept all other parameters equal. In order to do that, when
cutting the runs in order to obtain nseconds examples, we
only kept the first example, so that the number of generated
examples remained constant. We thus used data from the
interval [0, n]in order to perform the nseconds real time
estimation.
C. Impact of non the success rate and the mean error
We present in the Figure 3 the evolution of the success
rate (as always in generalization) and in the Figure 4 the
variation of the mean error in function of the value of n.
The success rate increases with the length of the interval,
while the mean error decreases. Increasing the window size
improves the quality of the examples, which indeed enhances
the results.
70
75
80
85
90
95
100
10 15 20 25 30 35 40 45 50
Success rate (%)
Window size (seconds)
Figure 3. Evolution of the success rate in function of the window size
0
0.05
0.1
0.15
0.2
10 15 20 25 30 35 40 45 50
Mean error(g/l)
Window size (seconds)
Figure 4. Network mean error as a function of the window size
successes failures success rate mean error
10s 104 28 78.79 0.155820
15s 102 30 77.27 0.143969
20s 108 24 81.82 0.142155
30s 111 21 84.09 0.123291
40s 110 22 83.33 0.125081
50s 116 16 87.88 0.105408
Table VI
REA L TIM E BL OOD A LC OHO L CO NTE NT E STI MATI ON O N A MIX ED S ET
(CO NTAI NIN G BOT H TR ACKS ), F OR A VALUE OF ǫO F 0.2 F OR
DI FFER EN T VALUE S OF T HE WI ND OW SI ZE (n).
D. Other real time results using all the available windows
The previously presented results focus on the impact
of the window size on the success rate and mean error.
Furthermore, we obtained similar results when using all the
available windows. Indeed, when we split our 60 seconds
long runs in windows of 20 seconds, we obtain 3 examples.
In the previous experiments we kept only the first one.
Keeping all the obtained examples gave similar results. This
also confirms the track independent estimations capability of
the network, since the windows [0-20], [20-40] and [40-60],
as an example, are 3 different sections of the track. Indeed,
only one lap is performed. Thus, the subject never drives
two times on the same section during a run.
We also obtained comparable results for the threshold ex-
ceeding detection (with better success rates), which we do
not present here as they are quite similar (slightly better)
than the ones detailed earlier in Section VIII-C.
IX. CO NC LU SI ON
We have demonstrated that it is possible not only to detect
threshold exceeding but also to estimate the blood alco-
hol content of a subject. These estimations were achieved
independently of the track. Considering that it is hardly
feasible to create an exhaustive learning base containing all
possible situations, it makes possible to only fill the learning
base with the most common situation patterns. We then rely
on the generalization capability of the neural network to
provide estimations on new patterns. We also managed to
obtain a real time estimation, and still independently of the
track played. The prototype is able to provide an estimation

using a 10 seconds window (e.g., after only 10 seconds of
playing), and to improve the reliability and the accuracy of
the estimation by using a larger window. It should thus be
possible to conceive an embedded interface that would be
able to provide estimations on the driver after only a short
while, and enhance these estimations after more time. The
fact that the estimations are not bound to a fixed length of
time allows many configurations.
Furthermore, such a system could re-evaluate the driver
continuously, and adapt the estimations to a changing situ-
ation. Indeed, someone who just drunk some alcohol may
feel perfectly able to drive, but as alcohol is absorbed in
blood, the side effects will alter his driving skills up to
the point where driving becomes dangerous. An embedded
device based on our concept could detect the evolution of the
driver and warn him before he becomes really dangerous.
X. LIM IT S AN D DR AWBACKS
While we could solve the problematics, we had to degrade
the accuracy of the system (ǫ) in some case in order to main-
tain acceptable success rates. As for track independence,
we note in some cases a significant drop on success rate
and an increase of the average error. But this experiment
was performed to see how the network would cope with
unknown tracks. Considering that we only had two tracks,
only one remained for training. As shown, if more accuracy
is required, we can use several tracks. Furthermore, if this
system were to be used in a production environment, we
would try to create a base containing a representative set of
the situations that can be encountered.
We also noted the much higher average error for real time
estimations when using small windows such as 10 s, but
it was expected. It indeed was intended to show the neural
network behavior in an extreme case, setting the lower bound
of the window size.
In some cases we reached an accuracy of 0.11 g/l, which
is close to the 0.1 g/l of the breathalyser. Owing to the
fact that we calibrated our system with this breathalyser,
improving the accuracy can only be achieved by using a
better breathalyser. On some cases though, the accuracy were
quite far from the breathalyser’s. As the present prototype
was created as a proof of concept, it was not intended to be
on par the breathalyser in every experiment.
XI. PE RS PE CT IV ES
We demonstrated the ability of the prototype to perform
meaningful estimations even in a simple environment. Those
results will be the base of our next works: we plan to
improve and test the system on a more realistic environment
in so that it might be embedded in a vehicle later on.
The prototype demonstrated the viability of the concept in
a simplified environment, we will now take it to a more
complex virtual world before trying to use it in the real
world. For now, much work is left in order to achieve this
goal: we will have to find what data to gather from a more
realistic system, or among the numerous data monitored
by modern cars, and also a way to accurately measure
those data. Considering the amount of completely different
situations in a real driving environment (the “game play” of
SuperTuxKart is very simplified), we will have to use either
much more inputs, or other sorts of neural networks such as
mixture of neural experts [4] to cut down the problematics
in smaller and more simple ones, considering how different
it is to drive on the highway and in congested urban areas.
No matter what neural network model we will use, it
will anyway imply much more work to select meaningful
inputs for the network, and probably require the use of
parallel computing in order to explore configurations until
we obtain the optimal ones for a given problem. Hopefully,
generalization will always be possible on a low cost, low
power consumption, and low computing power system,
making the software easily usable in embedded devices.
Furthermore, we foresee that increasing the complexity of
the environment and the amount of measured characteristics
might enable us to perform not only more accurate esti-
mations, but also many other estimations, such as detection
of tiredness, attention drop, or the use of driving impairing
drugs or medications. Those very similar problematics may
use the same system with just another examples set.
REF ER EN CE S
[1] Alcopass. Alcopass AL7000 breathalyzer documentation.,
December 2010. http://www.alcopass.com/.
[2] C. M. Bishop. Neural Networks for Pattern Recognition.
Clarendon Press, Oxford, 1995.
[3] E. Bogen. Drunkenness, a quantitative study of acute alcoholic
intoxication. California and Western Medecine, XXVI(6):778–
783, 1927.
[4] P. Estvez, H. Paugam-Moisy, D. Puzenat, and M. Ugarte. A
scalable parallel algorithm for training a hierarchical mixture
of neural experts. Parallel Computing, 28(6):861–891, 2002.
[5] J. Henrichs, M. Gagnon, and C. Pelikan. Super-
TuxKart, Free 3D kart racing game., December 2010.
http://supertuxkart.sourceforge.net.
[6] Ron Kohavi. A study of cross-validation and bootstrap for
accuracy estimation and model selection, 1995.
[7] Logitech. Formula Vibration Feedback Wheel., December
2010. http://www.logitech.com/en-in/441/298.
[8] S. Nissen. Implementation of a fast artificial neural network
library (FANN). Technical report, Department of Computer
Science University of Copenhagen, October 2003.
[9] D. Puzenat and I. Verlut. Behavior analysis through games
using artificial neural networks. In Third International Con-
ferences on Advances in Computer-Human Interactions, pages
134–138, February 2010.