Evolving a rule system controller for automatic driving in a car

racing competition
Diego Perez, Yago Saez Member, IEEE, Gustavo Recio, Pedro Isasi

Abstract The techniques and the technologies supporting

Automatic Vehicle Guidance are important issues. Automobile
manufacturers view automatic driving as a very interesting
product with motivating key features which allow improvement
of the car safety, reduction in emission or fuel consumption or
optimization of driver comfort during long journeys. Car racing
is an active research field where new advances in aerodynamics,
consumption and engine power are critical each season. Our
proposal is to research how evolutionary computation techniques can help in this field. For this work we have designed an
automatic controller that learns rules with a genetic algorithm.
This paper is a report of the results obtained by this controller
during the car racing competition held in Hong Kong during the
IEEE World Congress on Computational Intelligence (WCCI
2008).

I. I NTRODUCTION
In 1970, the visionary Robert E. Fenton, in his survey,
predicted how automatic vehicle guidance would evolve, [1]:
The system would probably be implemented in three overlapping stages. The first could be the installation and use
of various driver aids so that the driver would be a more
effective decision maker and improve the performance of
the driver-vehicle system. The second stage could involve
the gradual introduction of various subsystems for partial
automatic control. The third would include the transition
to complete automatic vehicle control. Each of these stages
must be realized within the confines of one system so that
the addition of each feature would contribute to the ultimate
system.
Today, we are in the second stage: we can buy cars with
electronic aided brake systems (ABS), lane change aids,
adaptive cruise controls which maintain a set distance from
the car ahead, automatically accelerating or decelerating,
and even applying the brakes, etc. All these new technical
advances are a good starting point for approximating the third
stage: to complete automatic vehicle control.
The techniques and the technologies supporting Automatic
Vehicle Guidance (AVG) are an important issue [2]. Automobile manufacturers view automatic driving as a very
interesting product with motivating key features which allow
improving of the car safety, reduction in emission or fuel
consumption or optimizing of the driver comfort during long
journeys.
This topic has been addressed by numerous researchers,
e.g. [3], [4], [5], [6], [7], [8], as an engineering
problem. However, this problem can also involve, among
The authors are with Carlos III University of Madrid, Av. de la
Universidad 30, Madrid, Spain (contact phone: +34-91-624-8456; email:
yago.saez@uc3m.es).

978-1-4244-2974-5/08/$25.00 2008 IEEE

others, robotics, artificial intelligence, computer engineering,

telecommunications, signal and image processing, or control
and automation techniques in order to develop efficient
systems for automatic driving.
Many different approaches have been studied in recent
years and the most promising ones are being engineered
on real prototypes, i.e., [6], [9] in the ARGO project, the
Buick from the California PATH project or [10]. This is due
to the fact that today more technological possibilities allow
development of completely functional prototypes with lower
costs.
In fact, there is a well-known annual competition organized by the Defense Advanced Research Projects Agency
(DARPA) for driverless cars which evaluates all the international improvements carried out annually in this field. An
example can be found in [11]. However, researchers have
a long way to go until these intelligent vehicles capable
of driving in a fully automated way are actually available.
Most of the works mentioned before was engineered in real
prototypes, this involves two main problems: the costs of
buying and modifying a car, and the time needed for each
test. To overcome these constraints we propose to use car
simulators. Today, car simulators are very close to reality,
and they allow us to speed up the research by testing more
techniques.
A. Evolutionary computation techniques applied to the automated driving
Our proposal for this task is to research how evolutionary
computation techniques can help in automatic driving. For
this proposal we have collected some work related to evolutionary computation techniques applied to automated driving.
These works fall into two different categories: autonomous
driving and vehicle features optimization.
1) Autonomous driving: one of the first approaches in this
area was carried out at the Carnegie Mellon University by
Sukthankar et al. in 1996, [7]. This work uses reasoning
modules which combine high level task goals with low-level
sensor constraints. Those modules are directly dependent
on a large number of parameters, and the setting of these
parameters must be done carefully. As this manual selection
is tedious and error prone, a Population Based Incremental
Learning (PBIL), which is a combination of Genetic Algorithms (GAs) and competitive learning, [12], is proposed for
automatically setting each modules parameters.
Therefore, the algorithm will propose what the probabilities of using a set of rules are depending on each situation.
The evaluation function takes into account different aspects,
336

such as serious crashes, collisions, wrong exits, distance

completed, etc. For the simulation, the system uses a program
called SHIVA (Simulated Highways for Intelligent Vehicle
Algorithms) which reproduces a microsimulation of vehicles
moving and interacting in a user-defined roadway. The algorithm can influence the vehicles motion sending simulated
commands (steering, accelerate and brake). In addition, the
system provides a perception module responsible for obstacle
detection, positioning, lane tracking, vehicle sensing, etc.
Two years later the PBIL was compared to the GA in this
same framework, [13], see Figure 1. The results of these
works were quite motivating; at the end of experiments the
vehicles automatically entered the test track, completed one
lap and a half and finally took the exit. Although they were
not capable of avoiding collisions with other vehicles, these
experiments showed the potential for intelligence behavior
in the tactical driving.

Fig. 1.

GA represents the vehicle movements with the starting and

ending points of the path, and a measure of the position with
the first and second derivatives of the path in a point near
the middle set of cones. The results were very good but they
only reached a near optimal solution due to the constraints
of the genetic representation selected.
In 2004, Floreano et al., [15], imitating strategies observed
in simple insects used a GA to tune up a neural network
which visually recognizes edges, corners and height. This
active vision system acts as an artificial retina, moving and
focusing on important features. It was tested with the open
source simulator Car-World (http://carworld.sourceforge.net)
and the best evolved individuals performed equal to or better
than well-trained human drivers tested on the same circuits.
In 2005, Sun et al. [16], used a GA to optimize the
parameters of a set of Gabor filters in the context of vehicle
detection from images. They tested the proposed framework
on real data with success and improved the performance of
on-road vehicle detection.
In the same year another interesting approach was proposed for automated evolutionary design of driving agents,
[17] [18] (Figure 2). This work showed how GAs can help
in the task of designing an agent able to remotely operate a
scale racing car. The agent perceives the environment from
a camera mounted overhead (position, orientation, velocity,
approach angle, distance to the apex and outside/inside slow
down zone). With these perceptions the agent sends commands to the remote control car (forward, neutral, reverse,
left, straight or right). The comparative analysis established
that on long runs the agents operated car was 5% slower
than the human operated one.

An example of the Cyclotron test track.

Other interesting work carried out in 1996 by Pyeatt et

al. from the Colorado University deals with simulated race
car driving [14]. In this case a study on autonomous driving
was developed based upon RARS simulator software (also
known as TORCS). This simulator gives information about
the vehicle position, vehicle speed, distance to the current
track segment, curvature of the track, and relative position
and velocity of nearby cars. It also offers the possibility of
controlling the speed and steering of the vehicle.
This work applied neural networks for learning automatic
driving through the use of input such as position, speed,
distance measured till the end of the segment, angle between
the vehicle and the road, etc. This neural network produced a
set of rules which decided when to accelerate, brake or steer.
The results showed that the RARS simulator was adequate
for developing the test framework and that neural networks
can be competitive techniques for producing autonomous
racing cars.
In 1998, Bernard et al. from the Iowa State University
illustrated the power of GAs to model driver/vehicle behavior. In fact, their work determined how fast and safe a
given vehicle model could be driven through a short course
without failure from hitting a cone or lifting the wheels. In
this work they used a simulator which they carried out. The

Fig. 2. Scheme of the system configuration for the 1:24 scaled model used
in [18].

Working with the evolving weights of a neural network,

Julian Togelius et al. compared, with their own simulator,
that simulated cars with evolved neural network controllers
(in first-person and third-person) [19], [20]. They extended
their work to a more complex case of two cars competing
against each other in the same track at the same time, [21],
using evolutionary strategies to solve the problem of the
co-evolution. Finally, an interesting study which compares
neuroevolution and genetic programming in the same environment can be found in [22].
2) Vehicle parameter optimization: the study of the suspension system has been an interesting topic for researchers
because it contributes to the cars handling and braking for
good active safety and driving pleasure, and keeps vehicle
occupants comfortable and reasonably well isolated from

2008 IEEE Symposium on Computational Intelligence and Games (CIG'08)

337

road noise, bumps, and vibrations.

In 1996 Cambiaghi et al. presented the results of a study
conducted on the vertical dynamics of a two wheel car
model [23]. They applied a GA to dampers optimization
with their own simulator. The results showed a significant
improvement in performance and comfort level compared to
the traditional damper behavior. Unfortunately no experimental results with real vehicles were available.
Another relevant study about GAs applied to suspension
systems is [24]. This work uses parallel GAs to optimize
three suspension key parameters: longitudinal center of gravity, roll stiffness distribution and aerodynamics downforce
distribution. The GA tests the parameters in a simulator
designed by the authors and promotes the fastest individuals.
The study helped in finding which values of the parameters
were the most adequate and what the relationship between
them was.
Finally, a recent work, [25], showed how to optimize 66
variables from a Formula one racecar with a GA. The parameters affect to the suspension system, the engine revolutions
limits and gear ratios, the aerodynamics, and the tyre and
brake modelling. These variables were tested in Electronic
Arts Formula One Challenge 99-02. This is a commercial
racing simulator released in 2003 with an advanced real time
physics engine. The results of this work were of interest,
the best car setup found involved performance improvements
(faster lap times) compared to the default system setup or
expert players settings.
II. O BJECTIVES
The objective of the developed controller is to drive as
good as possible in unknown circuits alone or in the presence
of other drivers. To this end we decided to participate in the
racing car competition held in the IEEE World Congress on
Computational Intelligence (WCCI 2008).
A. TORCS Simulator
The software used (TORCS: The Open Racing Car Simulator, source: http://torcs.sourceforge.net/) is today one of the
most popular simulators. This simulator is written in C++
and can be downloaded under GPL license from its web
page. The source code and the executable files are available
for Windows and Linux (for MacOS, only binaries are
accessible), which allowed the competitors more flexibility to
develop the controllers. The advantages of using this racing
simulator are the following:
TORCS has a high level of realism in graphics and
physics.
It provides a large quantity of vehicles, tracks and
controllers.
There is a large community of users and developers,
which helps the program to be updated and be kept to
a small number of errors.
It is easy to develop a controller and to integrate it
within the simulator.
However, one of the main problems of this simulator is
a memory leak each time the race is restarted. This is an
338

important disadvantage when the learning algorithm requires

a large number of iterations or evaluations, where each one of
them needs the race to be restarted. For instance, when a GA
needs to evaluate a whole population of individuals, it will
restart the race for each one of them, causing an increment
of this memory leak and, finally, inducing the program to
crash.
For this competition, the simulator has been adapted to
offer a common interface to all competitors. A server bot
is executed by the simulator, while each participant tests
his controller as clients. The communication among them
is designed using UDP packages, so each competitor can
develop his controller by choosing his favourite programming
language, with the sole condition of obeying the communication protocol.
B. Controllers, Sensors and Effectors
The interface provided by the server bot offers the client
a total of 17 sensors and 5 effectors. The sensors include
different kinds of information such as the angle of the car, the
position, speed, location of opponents, etc. The effectors used
to drive the vehicle are the steering wheel value, both pedals
(throttle and brake) and gearing change. A full description
of all the sensors provided can be seen in the bases of
the competition (documentation, sources and examples are
available at http://cig.dei.polimi.it/?page id=5).
Five controllers were presented to the competition and
three more were provided by the organizers. The aim of the
programmed controllers is to make a comparison against the
results obtained by the automatically developed ones.
C. Controller evaluation
The evaluation of the controllers results from their performance on three different test circuits made by the organizers.
The tests are done in different tracks which are previously
unknown by the competitors and divide the tournament into
two stages. At the first stage, every controller is thrown alone
in each circuit, evaluating the distance raced in 10, 000 game
tics. To obtain trustable results, ten executions are performed
on each track, considering the mean of all of them for point
assessing. In the second stage, all controllers (except the three
programmed by the organizers) are tested together, again ten
times, measuring the arrival order to evaluate the drivers.
In both cases, the points assigned to each controller
follows the Formula One World Championship pointscoring
system (10 points for the first, 8 for the second, 6 for the
third, and so on, up to one point for the eighth).
III. C ONTROLLER D ESIGN
The design of the controller is explained in this section,
including a description of problems found and decisions
taken through the development.
A. Input Data
Among all the available sensors, only four of them have
been used to create the rule system. These input data were

2008 IEEE Symposium on Computational Intelligence and Games (CIG'08)

discretized in order to obtain a set of rules used by the vehicle

to drive in any circuit. The sensors used are the following:
Angle (Figure 3): angle between the car direction and
the direction of the track axis. The target range is [0,4],
where 0 means small angles and higher values represent
larger ones.

front one and the immediate sensors on its right and

left) and they have been discretized to a unique range
[0,2]. In this case, 0 means that the track edge has been
detected between 20 and 100 meters, 1 means that the
track edge is up to 20 meters and 2 that no track edge
is seen.
B. Effectors
The effectors of the controller have been designed as
follows:
Throttle and brake: both pedals have been codified in
a common output, in order to avoid non-sense values.
For instance, a full gas value in both pedals at the same
time would produce uncertain (and useless) outcomes.
Hence, a unique value (a) is applied and both gas pressures are extracted from it using the formula depicted
in the Figure 6.

Fig. 3.

Vehicle angle with track axis discretization.

TrackPos (Figure 4): distance between the car and the

track axis. This continuous value has been discretized
in a discrete range [0,1], where 0 means centered in the
track and 1 represents the car near the edges.

Fig. 6.

Fig. 4.

Vehicle track position discretization.

An important feature of the design is the usage of symmetry for the first two sensors. This concept works by using the
absolute value of the sensor to match to the proper discretized
value. In other words, a value of and from any of those
sensors will mean the same discretized value. The objective
of this approach is two-fold: to reduce the search space and
to avoid the algorithm learning how to face similar situations
twice, taking into account that the difference is only the sign
of the sensor value.
SpeedX (Figure 5): speed of the car along the longitudinal axis of the car, discretized in a range [0,3], where
0 means lower speed than higher values.

Fig. 5.

Vehicle speed discretization.

Track: the set of sensors that indicates the distance to

the track edges. Only three of them have been used (the

Formula for pedals values.

Steer: the steering output is codified as a real number,

taking the same values as the effector, from -1 to 1, but
discretized with a precision of 0.1. The reason for this
discretization is to reduce the number of possibilities,
considering only the ones with most relevance: notice
that a difference of 0.001, for instance, in steering
does not become in an important change on how the
car steers.
Gear: gear changes did not take part in the learning
process, and have been defined as follows: if the revolutions per minute of the engine are higher than 6, 000,
then the gear value is increased by one. However, if this
value is lower than 3, 000, the current gear is decreased.
Otherwise, the gear does not change.

C. Rules
The discretization applied over the input data allows us to
create a set of 120 rules (Figure 7), where the conditional part
is composed of the above four sensors used and the actions
are formed by acceleration, braking and steering, which are
codified as seen previously. This set of initial rules are the
basis of the base individual.
D. Base individual
One of the biggest problems found when evolutionary
techniques are applied to obtain autonomous driving is
that traditional random initialization of individuals does not

2008 IEEE Symposium on Computational Intelligence and Games (CIG'08)

339

Fig. 7.

Set of rules scheme.

work properly because it is almost impossible to obtain a

configuration that drives the vehicle correctly by chance. This
is the reason for getting a base individual capable of finishing
at least one lap in the circuit, and then evolve it to obtain
better results.
The algorithm used to obtain this base individual is the
generation of a subset of rules that allows the vehicle to
end a lap, minimizing the angle of the car with the track
axis, no matter the speed of the vehicle. Each one of these
rules is created by testing how each allowed combination of
acceleration and steering behaves when the left part of the
rule (condition) is triggered. The result of this procedure is
a novice controller who drives at a low speed and always
centered on the track.
E. Evolutionary Rule System
Once the base individual is obtained, an evolved individual
is generated from it, taking all the rules used in the previous
process. Obviously, not every one of the set of 120 rules is
used, so the final individual will only use a subset of them
(results show that this number is usually between 10 and 20),
depending on the circuit used to obtain the base individual.
Therefore, the individual is composed of a set of rules, each
one of them formed by the sensors condition and the effectors
action, that need to be evolved to obtain the controller. An
scheme of this individual is depicted in Figure 8.

Fig. 8.

training track, but to obtain a vehicle capable of taking the

turns correctly. On one hand, if the only factor to take into
account in the evaluation of the individual were the lap times,
then the vehicle would learn how to handle along the bends
of this specific circuit, as did in [26], and it would be useless
for other tracks. On the other hand, if the car suffers a lot of
damage, it will end up stopped somewhere in the circuit (and
that is negative if we wish to drive many meters, as requested
in stage one of the competition). Thus, this parameter was
then included in the fitness evaluation.
In this rule system, we can not decide when a rule is
better than the other because the behaviour of the individual
depends on the whole set of rules used. Because of this,
a selection operator has been designed as a random pick-up
from the pool of rules, taking two of them to apply crossover.
This operator has been implemented as an uniform crossover,
using a probability of 50% to choose a gene from each parent
rule. Finally, a mutation is performed over the new rule,
applying an addition of 1 unit to the left part and 0.3
to the effectors of the rule (always obeying the limits of the
genes and their codification precision, as was specified in
previous sections).
The next step in this algorithm consists of the search of
the rule in the individual with identical conditional part to
the new one. This rule is extracted from the pool and the new
one replaces it. Notice that the impact of this substitution is
not too high because of the similarity protocol. After that, the
new set of rules is evaluated, and the fitness is compared with
the one calculated before inserting the new rule until a stop
criterion is met. If the new set obtains a better fitness, the
new rule stays in the individual. Otherwise, the substituted
rule is retrieved and the new one is eliminated. In any case,
a new step of the algorithm is executed until a stop criterion
is reached.
Each one of the steps of this algorithm is shown in the
scheme of Figure 9.

Base individual scheme.

Fig. 9.

To evolve this individual, a circuit is chosen (complex

enough to provide multiple and different driving scenarios)
and the following algorithm is executed:
The evaluation of the individual is performed by recording
the lap time and damage suffered when it is executed in
a circuit for three laps. The fitness assigned to the set of
rules is obtained through a lineal combination of both values,
with weights of 0.4 and 0.6 respectively. This adjustment is
made in order to avoid the overfitting to the training circuit.
The objective here is not to obtain the best lap time on the
340

Algorithm step.

Results shows that this algorithm is effective since it reduces the base individuals lap times in few generations, and
keeps the damage of the car almost inexistent. However, more
tests need to be done in order to extract some conclusions.
IV. R ESULTS
The controller developed for the competition (called
DIEGO) was tested in three circuits and it obtained acceptable results in the first two (Figures 10 and 11).

2008 IEEE Symposium on Computational Intelligence and Games (CIG'08)

TABLE I
F INAL RESULTS .

Fig. 10.

Ruudskogen circuit results at Stage 1.

Fig. 11.

Street1 circuit results at Stage 1.

After studying the results we concluded that the usage of

symmetry produced a side effect that was not expected: the
car drives in a smooth zig-zag trajectory centered on the
track. This is because a small steering value can center the
vehicle on the track, but not necessarily drive it parallel to the
track axis, because symmetry changes the sign of the steering
value when the track axis is crossed. However, the controller
behaved in a reasonable way, only with just the exception of
some specific circuits: the oval ones. These circuits, similar
to Nascar tracks, have banked bends which make the zig-zag
movement completely uncontrollable.
This problem affected dramatically to the behaviour of this
driver in the third circuit of the test bed, as can be seen in
the results shown in Figure 12.

Fig. 12.

D-Speedway circuit results at Stage 1.

In the second stage, the races performed again over these

circuits gave the last set of points assigned to the drivers. The
final results can be seen in Table I. Note that the manually
programmed controllers are not included (LUIGI, DANIELE
and JULIAN)

The results of the races are very similar to the results

obtained in the first two circuits on stage 1, far from the
best controllers, but with reasonable positions. However, the
results obtained in the third circuit on stage 1 weighted the
final score in such a way that finishing in higher positions
became impossible.
As seen before, the main difference between the DSpeedway circuit and the others in the bed test was the
presence of banked bends. For that reason, another attempt,
which considered an oval circuit (B-Speedway) with these
kind of banked bends for the training process, was made
after the competition so the vehicle could adapt better to
these conditions.
After these modifications, the experiment results for the
three tracks of the competition yielded: 5, 297 meters on
Ruudskogen, 4, 499 meters on Street1 and 5, 670 meters on
D-Speedway. These results cannot be directly compared with
those obtained by all other competitors due to hardware
factors, technical features of the computer used for the
purpose of this work and the one used for the competition are
different, thus, some variations are expected. However, these
results showed that the driving of the vehicle improved using
this kind of training circuits. After obtaining these results
we expect great improvements in the future controllers when
training them with more circuits.
V. F UTURE W ORK
Considering the limited time available to prepare the
controller for the competition, we can evaluate our work
as positive, since we only failed in one of the test circuits.
However, the main aim now is to continue working on this
driver in order to obtain a very good controller.
Competition organizers decided to propose a new edition
of this tournament for the IEEE Symposium on Computational Intelligence and Games (CIG2008), so the main
objective now is to improve this controller in order to obtain
better results for this new call. The next steps to follow are:
Discretization: the input data discretization was performed by hand, without using any methodology that
help us to determine the target ranges. The usage of a
technique of clustering (for instance, applying k-means
over human driving in the simulator) could determine
more representative ranges.
Symmetry: as has been previously said, symmetry
caused a big problem in some circuits, despite the fact
that it has decreased the search space and helped the
evolutionary algorithm to obtain better results. The next

2008 IEEE Symposium on Computational Intelligence and Games (CIG'08)

341

controller will introduce a system without symmetry, or

using a limited symmetry, to avoid the zig-zag movement of the vehicle. If we finally get better discretization
ranges, as stated in the last point, the impact on the
search space would not be a problem for the final design.
Evaluation in more than one track: TORCS simulator
does not easily allow restarting the race in a circuit
different than the one used for starting the execution.
This causes the learning to be performed in only one
circuit, which can produce overfitting on the training
track, despite the attempts to choose a complex circuit to
force the vehicle to face a set of different scenarios (such
as bends, track widths, circuit lengths, etc.). The idea
is to use more than one circuit in the learning process
collecting more efficient rules for the controller in all
circumstances.
Opponent analysis: an interesting point observed in
race competition videos is that most of the delivered
controllers can not avoid the opponents when they are
about to overtake the other cars. Instead of that, the
vehicles hit them in the back without evading them. For
the next competition, an analysis of these opponents will
be made in order to obtain a competitive advantage in
the race phase.

ACKNOWLEDGMENTS
This work was supported in part by the Spanish MCyT
project OPLINK, under grant no. TIN2005-08818-C04-02.
R EFERENCES
[1] R.E. Fenton. IEEE Transactions on Vehicular Technology, 19(1):153
161, 1970.
[2] J. Bernard, J. Gruening, and K. Hoffmeister. Evaluation of vehicle/driver performance using genetic algorithms. SAE International
Congress and Exposition, (980227), 1998.
[3] A. Niehaus and R. F. Stengel. Probability-based decision making
for automated highway driving. Vehicle Navigation and Information
Systems Conference, 1991, 2:11251136, Oct. 1991.
[4] G. Siegle, J. Geisler, F. Laubenstein, H. Nagel, and G. Struck.
Autonomous driving on a road network. Intelligent Vehicles 92
Symposium., Proceedings of the, pages 403408, Jun-1 Jul 1992.
[5] R. Sukthankar, D. Pomerleau, and C. Thorpe. Shiva: Simulated
highways for intelligent vehicle algorithms. In In Proceedings of IEEE
Intelligent Vehicles, pages 332337, 1995.
[6] D. Pomerleau and T. Jochem. Rapidly adapting machine vision for
automated vehicle steering. IEEE Expert, 11(2):1927, Apr 1996.
[7] R. Sukthankar, J. Hancock, S. Baluja, D. Pomerleau, and C. Thorpe.
Abstract adaptive intelligent vehicle modules for tactical driving. In In
Proceedings of AAAI-1996 Workshop on Intelligent Adaptive Agents.
[8] C. Thorpe, T. Jochem, and D. Pomerleau. Automated highway and
the free agent demonstration. In In Proceedings of 1997 IEEE Conf.
on Intelligent Transportation Systems, pages 496501, 1997.

342

[9] M. Bertozzi, A. Broggi, G. Conte, and R. Fascioli. The experience of

the argo autonomous vehicle. In in Procs. SPIE98 - Aerosense Conf,
volume 3364, pages 218229, 1998.
[10] J. M. Collado, C. Hilario, A. de la Escalera, and J. M. Armingol. Selfcalibration of an on-board stereo-vision system for driver assistance
systems. Intelligent Vehicles Symposium, 2006 IEEE, pages 156162,
June 2006.
[11] Q. Chen, U. Ozguner, and K. Redmill. Ohio state university at the 2004
darpa grand challenge: developing a completely autonomous vehicle.
Intelligent Systems, IEEE, 19(5):811, Sept.-Oct. 2004.
[12] S. B. and R. Caruana. Removing the genetics from the standard genetic
algorithm. In The Int. Conf. on Machine Learning 1995, pages 3846,
San Mateo, CA, 1995. Morgan Kaufmann Publishers.
[13] S. Baluja, R. Sukthankar, and J. Hancock. Prototyping intelligent
vehicle modules using evolutionary algorithms. pages 241 257, 1998.
[14] L. D. Pyeatt, A. E. Howe, and C. W. Anderson. Learning coordinated
behaviors for control of a simulated robot. Technical report, Computer
Science Dept, Colorado State Univ., Ft. Collins, CO 80523, 1996.
[15] D. Floreano, T. Kato, D. Marocco, and E. Sauser. Coevolution of
active vision and feature selection. Biological Cybernetics, 2004.
[16] Z. Sun, G. Bebis, and R. Miller. On-road vehicle detection using evolutionary gabor filter optimization. IEEE Transactions on Intelligent
Transportation Systems, (6):125137, 2005.
[17] I. Tanev, M. Joachimczak, H. Hemmi, and K. Shimohara. Evolution
of the driving styles of anticipatory agent remotely operating a scaled
model of racing car. Evolutionary Computation, 2005. The 2005 IEEE
Congress on, 2:18911898 Vol. 2, Sept. 2005.
[18] I. Tanev, M. Joachimczak, and K. Shimohara. Evolution of driving
agent, remotely operating a scale model of a car with obstacle avoidance capabilities. In Mike Cattolico, editor, GECCO 06: Proceedings
of the 8th annual conference on Genetic and evolutionary computation,
pages 17851792, New York, NY, USA, 2006. ACM.
[19] J. Togelius and S. M. Lucas. Evolving controllers for simulated car
racing. Evolutionary Computation, 2005. The 2005 IEEE Congress
on, 2:19061913 Vol. 2, Sept. 2005.
[20] J. Togelius and S.M. Lucas. Evolving robust and specialized car racing
skills. Evolutionary Computation, 2006. CEC 2006. IEEE Congress
on, pages 11871194, 2006.
[21] J. Togelius and S. M. Lucas. Arms races and car races. In Parallel
Problem Solving from Nature - PPSN IX, 9th International Conference,
Reykjavik, Iceland, September 9-13, 2006, Procedings, volume 4193
of Lecture Notes in Computer Science, pages 613622, 2006.
[22] A. Agapitos, J. Togelius, and S. M. Lucas. Evolving controllers for
simulated car racing using object oriented genetic programming. In
GECCO 07: Proceedings of the 9th annual conference on Genetic and
evolutionary computation, volume 2, pages 15431550. ACM Press,
2007.
[23] D. Cambiaghi, M. Gadola, L. Manzo, and D. Vetturi. Semi-active
strategies for racing car suspension control. SAE International
Congress and Exposition, (962553), 1996.
[24] E. M. Kasprzak, K. Hacker, and K. Lewis. Exploring the design
tradeoffs and computational savings of executing vehicle simulations
in a parallel computing environment. In ASME Design Automation
Conference, DETC00/DAC-14243, 2000.
[25] K. Wloch and P. J. Bentley. Optimising the performance of a formula
one car using a genetic algorithm. In Parallel Problem Solving from
Nature - PPSN VIII, 8th International Conference, Birmingham, UK,
September 18-22, 2004, Proceedings, volume 3242 of Lecture Notes
in Computer Science, pages 702711, 2004.
[26] Y. Saez, D. Perez, O. Sanjuan, and P. Isasi. Driving cars by means of
genetic algorithms. In PPSN, pages 11011110, 2008.

2008 IEEE Symposium on Computational Intelligence and Games (CIG'08)