Development of a Simulation Environment for Automated Model

Configuration for the Design and Validation of AI-based Driving

Functions

Or Aviv Yarom

and Xiaobo Liu-Henke

Ostfalia University of Applied Sciences, Salzdahulmer Str. 46/48, 38302 Wolfenbuettel, Germany

Keywords: Autonomous Driving, Artificial Neural Networks, Matlab, Simulink, Deep Learning Toolbox.

Abstract: The further development of autonomous driving requires the increased use of innovative and intelligent

algorithms. In order to develop these effectively and efficiently, suitable development methods and tools are

required. Therefore, this paper presents the development of a simulation environment for automated model

configuration for the design and validation of AI-based driving functions. Based on the current state of the

art, the conception including requirement definition and realization of the simulation environment are

described in detail. In addition, the simulation environment is validated in an application for automated vehicle

guidance with Artificial Neural Networks.

1 INTRODUCTION

The innovation alliance autoMoVe (Dynamically

Configurable Vehicle Concepts for Use-Specific

Autonomous Driving), funded by the European

Regional Development Fund (ERDF), aims to

develop an autonomous, modular and electric vehicle

concept. By exchanging application specific modules

during runtime, a wide range of applications from

internal freight transport to passenger conveyance in

road traffic shall be realized autonomously. Within

the scope of this research project, the Ostfalia

subproject autoEVM (Holistic Electronic Vehicle

Management for Autonomous Electric Vehicles)

focuses on the model-based development of

innovative intelligent algorithms and functions for

autonomous driving.

Higher automation of driving operations is

accompanied by an increase in the requirements to be

met by the vehicle or the automated driving functions.

Current functions and algorithms based on methods

of control theory or classical information processing

can no longer fully meet these (Milz and Schrepfer,

2020). Therefore, artificial intelligence (AI)

represents a key technology in this project or for

many domains involved in the development and use

https://orcid.org/0000-0001-5627-4199

of intelligent, automated vehicles (Fayjie et. al.,

2018).

Regardless of the type of information processing,

modern vehicles and driving functions are complex

mechatronic systems with a high degree of internal

and external interconnection. In order to handle this

complexity in the development and validation

process, a design methodology that is well established

in mechatronics research is used. This consistent and

verification oriented methodology is based on digital

models and simulations to make the design and

validation process of complex mechatronic systems

in a crosslinked environment easier, faster and safer.

(Liu-Henke et. al., 2016)

The development and simulation environments

currently available for intelligent vehicle functions

are very extensive in general, but they concentrate

largely on conventional algorithms for information

processing. The use of AI functions in development

and validation is either not possible or only possible

with a great amount of effort. Conversely, current

development environments for AI algorithms do not

offer the advantages or the functional scope of tools

that are specifically designed for automated driving

functions. (Stančin and Jović, 2019)

This current incompatibility of the two worlds for

automated driving functions and AI is therefore now

Yarom, O. and Liu-Henke, X.

Development of a Simulation Environment for Automated Model Conﬁguration for the Design and Validation of AI-based Driving Functions.

DOI: 10.5220/0010611901630171

In Proceedings of the 11th International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH 2021), pages 163-171

ISBN: 978-989-758-528-9

163

inhibiting the research and development progress of

automated and autonomous driving. Consequently, in

order to realize the mobility transformation

envisioned in the projects and society, new methods

and tools are needed that can unite both worlds.

Therefore, this paper presents the development of a

simulation environment for automated model

configuration for the design and validation of AI-

based driving functions. On the one hand, this

simulation environment's scope of functions is geared

towards the development of automated driving

functions and, on the other hand, it offers the

possibility to use not only conventional but also AI

algorithms.

2 METHODOLOGY

The complexity of modern vehicles is constantly

increasing due to the higher degree of internal and

external networking and the growing number of

intelligent and efficient hardware and software

components. In order to handle the system

complexity and to avoid errors at an early stage in the

design of the information processing, a holistic design

methodology is indispensable. Therefore, the

continuous, verification-oriented, model-based

design methodology based on Rapid Control

Prototyping (RCP) and Model-in-the-Loop (MiL),

Software-in-the-Loop (SiL) and Hardware-in-the-

Loop (HiL) simulations has been established. (Liu-

Henke et. al., 2016)

The methodology is based on function-oriented

physical models of a controlled system. The control

function is then simulated depending on the system

behavior and validated in MiL simulations at an early

stage. To avoid manual programming, the model and

control function are developed in block diagram-

based programming languages. The subsequent

automatically generated function code is again tested

against the control system model in SiL simulations.

HiL simulations are used for further validation and

optimization of the information processing with real-

time capable simulation models and real

subcomponents of the system to be controlled.

The verification oriented and iterative approach

of this methodology also supports the development

process in the challenging task of validation. The

methodology addresses the weaknesses of classical

validation based on physical prototypes, such as a

high expenditure of resources or safety risks for

humans, machines and the environment. Due to their

virtual character, MiL, SiL and HiL simulations save

time and costs (Yarom et. al., 2020a). They enable

feasible and reproducible tests at any time without

direct dependence on physical prototypes, times of

day or human experts. Thus, simulation cycles, for

different functional variants or scenarios, can be

automated. This makes this methodology particularly

suitable for training AI algorithms. This is because,

with rare exceptions, machine learning is always

iterative.

Virtual design methods like these form the basis

for many intelligent systems, such as highly

automated vehicles. With prototype-based testing, the

hundreds of thousands of test kilometers required

would not be achievable in a reasonable amount of

time and at a reasonable cost. (Yarom et. al., 2020a)

3 STATE OF THE ART

3.1 Intelligent Driving Functions

In automated driving, individual driving tasks are

taken over from the human driver by so called

advanced driver assistance systems (ADAS). Such

ADAS, e.g. for speed control or lane keeping

assistance, have been available in series production

for some time (Kukkala et. al., 2018). ADAS process

the data collected by vehicle and environment sensors

and thus calculate driving commands, which are then

implemented by means of controlled actuators. With

an increasing number and interconnection of ADAS,

the human driver successively delegates driving tasks

to the vehicle until he finally becomes a passenger in

autonomous driving. Then we no longer speak of

ADAS, but of (automated) driving functions.

Increasing automation of the driving process

means an extreme increase in the requirements for the

driving functions. Not only more but also different

sensors are needed for environment perception,

whose inhomogeneous raw data must be processed

and implemented repeatedly, intensively and with the

highest real-time requirements. If information from

internal bus or external vehicle-to-everything (V2X)

communication is added, the complexity increases

even further. This pushes conventional algorithms for

control and data processing to their limits. (Milz and

Schrepfer, 2020) For this reason, AI algorithms are

already being used today for automated driving

functions. Artificial neural networks (ANNs) in

combination with machine learning (ML) are

particularly promising. Prominent applications are

image-based semantic segmentation of the driving

environment (Lyu et. al., 2019) or automated vehicle

guidance (Huang et. al., 2019). AI algorithms are

promising for the further development of autonomous

SIMULTECH 2021 - 11th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

164

vehicles due to their performance, robustness and

adaptability (Kuutti et. al., 2021). There is also

potential for intelligent vehicle functions outside of

automated vehicle guidance, e.g., for battery

management (Alaoui, 2019). On the other hand,

conventional approaches are often still used at lower

levels of information processing, e.g. local control of

actuators (Milz and Schrepfer, 2020).

3.2 Fundamentals of Artificial Neural

Networks and Machine Learning

The term AI covers a variety of different methods and

algorithms that deal with the autonomous and

automated solving of problems (Togelius et. al.,

2018). ANNs and ML form a subfield of AI that has

been shown to be suitable in numerous problems in a

wide variety of domains, including autonomous

driving. Therefore, this paper focuses on this subfield.

The numerous positive properties of ANNs and ML,

such as adaptability, error resistance, versatility, and

above all learning ability, can be traced back to their

similarity to the structure and functioning of the

human brain.

Analogous to biology, (artificial) neurons are

processing units that accumulate input stimuli via

weighted connections and compute an output using an

activation function. The interconnection of several

neurons in at least two layers makes up the ANN.

Combinations of up to several hundreds of neurons in

up to more than one hundred layers are common. Not

only arbitrary forward but also time-feedback

connections are possible in the ANN. (Skansi, 2018)

The optimal architecture of an ANN cannot be

determined analytically so far (Tirumala, 2020).

Therefore, experience and test series are necessary to

find a suitable architecture in the trade-off between

computational effort and performance. The number,

interconnection and weighting of connections

characterizes the "intelligence" of an ANN. Generally

speaking, more neurons and connections mean a

higher performance of the ANN, while at the same

time the computational effort increases.

Just like a human brain, the ANN must first learn

or train a task. These terms refer to the adaptation of

the connection weights. In the environment of

autonomous driving, supervised learning (SL) and

reinforcement learning (RL) are relevant for this. In

SL, the ANN is provided with input data and the

corresponding output. The ANN iteratively learns the

relationship between the two variables (Duriez,

Brunton and Noack, 2017). This learning procedure

is particularly suitable for image based object

recognition, for example (Lyu et. al., 2019). In RL,

the ANN successively learns the optimal strategy

from the experience of past sequences in terms of a

given reward function (Duriez, Brunton and Noack,

2017). This procedure is used when no training data

is available, e.g., in automated vehicle guidance

(Huang et. al., 2019). SL and RL are head categories

of learning procedures, with diverse concrete training

algorithms. Just like the ANN architecture, the

optimal training algorithms or their parameters

cannot be determined analytically. Thus, experience

and experimentation are required here as well.

3.3 Development Environments for

Intelligent Driving Functions

For the model-based design of automated driving

functions, several development and simulation

environments exist, such as MATLAB/Simulink with

toolboxes, dSPACE Automotive Simulation Models

(ASM), and IPG CarMaker, to name just a few. They

all differ in terms of their primary focus or their

specific advantages and disadvantages. All of them

offer extensive model libraries for traffic, vehicle

dynamics, component, sensor or even driver models.

They are specifically designed for the configuration

and reproducible simulation of a wide range of

driving scenarios and vehicle variabilities for the

purpose of vehicle development or validation. The

tools usually have integrated model configurators,

visualization, experiment environments, and in some

cases scenario and test managements. The tools listed

are suitable for the design methodology described in

section 2 using the RCP process. They enable not

only MiL simulations, but also real-time SiL and HiL

simulations through automatic code generation in

conjunction with MATLAB/Simulink. (Deter et. al.,

2021) However, the tools do not offer the possibility

to integrate ANNs into the respective simulation and

configuration tools without further effort, let alone to

train them.

The only exception here is MATLAB's Deep

Learning Toolbox (DLT). With this toolbox, any

ANN architecture can be configured and trained with

a large number of pre-implemented algorithms. The

DLT is basically compatible with Simulink and

automatic code generation. However, ANN

configuration is quite complicated and is usually done

manually for individual ANNs. As a result, it is not

very suitable in its native form for carrying out test

series with varying architectures, training parameters

or even different driving functions.

There are also specific development

environments for ANN training. Most of them are

based on the Python programming language.

Development of a Simulation Environment for Automated Model Conﬁguration for the Design and Validation of AI-based Driving Functions

165

Prominent tools are, among others, TensorFlow,

PyTorch, Keras and Caffe. They all have comfortable

and extensive functionalities regarding the creation

and training of ANNs. However, there is only the

possibility of scripted programming. A clearer block

diagram-based programming suitable for the RCP

process is not available. (Stančin and Jović, 2019)

Therefore, it is difficult to design simulations with the

same accuracy, computation time and comfort with

the ANN-specific tools as with the tools for

automated driving functions. The effort for this is so

large that even computer games like Grand Theft

Auto (Wang et. al., 2019) or TORCS (Zhang and Cho,

2017) have been coupled for vehicle and traffic

simulations in combination with the mentioned ANN

tools. However, these approaches rather served the

investigation of ANNs and ML and do not meet the

requirements of accuracy, real-time capability,

reproducibility and variation possibility of a real

driving function development.

4 CONCEPTION OF THE

SIMULATION ENVIRONMENT

4.1 Deriving the Problem Statement

The findings from section 3 can be used to derive a

problem statement for pursuing the project goals

(section 2). The design of intelligent automated

driving functions requires the use of AI algorithms.

However, as described, these cannot easily be used in

the usual development environments. ANNs and ML

methods quickly become very complex and

confusing. A manual programming of different ANN

architectures with the associated calculation rules

would not only be error-prone but would also take a

lot of time. The same applies to the ML procedures,

which would have to be parameterized and adapted

for each ANN architecture. The fact that architecture

design, training and test processes are always

empirical and iterative requires many simulation

cycles and further worsens the situation.

The use of ANN-specific development

environments is also ruled out. If the automated

driving functions are to be designed under realistic

conditions, the depth of modeling must be sufficiently

precise. In the classical way of analytical physical

modeling, the mathematical equations and their

numerical solution methods would have to be

programmed independently in a script language. If

one now considers the number of subsystems of a

vehicle, further driving functions or other road users

that are to be simulated, a highly complex simulation

system results. In addition, there are different

variants, configurations, scenarios and, if necessary,

real-time requirements. Although the implementation

of such a project would be possible in principle, it

would be very error-prone and not very effective.

Finally, there are already resource-optimal simulation

environments for this purpose.

In summary, none of the currently available

simulation environments is suitable for the design and

validation of AI-based driving functions according to

the development methodology from section 2. The

conclusion is therefore that a separate simulation

environment must be developed for this design

methodology.

4.2 Definition of Requirements

To address the aforementioned challenges, the

requirements for the new simulation environment for

automated model configuration for the design and

validation of AI-based driving functions are defined

below:

R1 Usable for various driving functions

R2 Operation with Simulink and compatibility

with corresponding blocks and models

R3 Use of various existing or creation of own

models and functions with any modeling depth

R4 Easy creation and calculation of ANNs

R5 Automatic integration of ANNs into the

Simulink models

R6 Variation of model configuration

R7 Automatic model configuration

R8 Automated execution of simulation series

R9 Training by means of different ML methods

R10 Use for generation of training data for ML

R11 Visualization in 2D and 3D

R12 Operation with user interface or scripts

R13 High-performance computing times for large

simulation series and later real-time

applications

R14 Compatibility with dSPACE ASM for later

development of the simulation environment

R15 Possibility of automatic code generation

4.3 Concept Formation

The first step to fulfill the previously defined

requirements, is the realization of the simulation

environment in MATLAB and Simulink (R2). This

ensures compatibility with standard blocks and

SIMULTECH 2021 - 11th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

166

existing models available in Simulink (R3).

Furthermore, any other models and automated driving

functions can be created (R1). In principle, toolboxes

can also be used for this purpose. It is only important

to ensure that these support automatic code

generation (R15). In this case, compatibility with

dSPACE ASM is also guaranteed (R14).

For the configuration and creation of ANNs, the

DLT is used. For this purpose, an additional ANN

generator is designed, which creates arbitrary ANNs

architectures automatically by simple user input via a

graphical user interface (GUI) or a script (R4) and

generates them directly into the simulation model

(R5). By using the DLT, several ML algorithms are

automatically available (R9). Several algorithms are

already usable for SL, given that one has generated

training data (R10). RL algorithms are available in the

RL toolbox of MATLAB. Alternatively, it is possible

to implement custom training algorithms for both SL

and RL, since the DTL provides convenient access to

the connection weights.

With a scenario generator, different simulation

scenarios can be configured with respect to the route

and the road participants (R6). This configuration is

then automatically transferred to the simulation

model. In order to better interpret the events in the

scenario during and after the simulation, an additional

rudimentary visualization is implemented. The 2D

and 3D visualization is based on the Bird's-Eye Scope

and the plot function of MATLAB (R11).

The configuration and creation of the scenario is

also done by a GUI or a script (R12). With the

possibility of script-based scenario and ANN

generation, simulation sequences can be automated

(R8). The corresponding model parameters are

automatically configured and updated (R7). Just as

with the developed driving functions and models, a

lean and computationally optimized programming

must also be taken into account for the realization of

the scenario and ANN generator as well as the

automation mechanisms (R13).

5 REALIZATION OF THE

SIMULATION ENVIRONMENT

Figure 1 shows the setup and structure of the

simulation environment. It mainly consists of the

simulation model itself and the control of the

simulation environment. In the simulation model, the

modeling and the calculation of the simulation take

place. It contains the models, parameters and

functions of the entire ego vehicle in the desired

Figure 1: Setup and structure of the simulation environment

for automated model configuration for the design and

validation of AI-based driving functions.

modeling depth as well as the environment simulation

defined in the scenario. The ego vehicle is the "own"

vehicle for which the automated driving functions are

developed. The ego vehicle is usually modeled in a

very detailed way (in contrast to the other third-party

vehicles) and contains several components in the

desired modeling depth, depending on the

application. These include, for example, the driving

functions to be developed (AI-based or otherwise),

sensors, vehicle dynamics, communication or

component models. These individual model

components of the ego vehicle communicate with

each other as well as with scenario model components

via defined interfaces. The scenario model includes

the various components of the driving environment,

such as route models with infrastructure, third-party

vehicles, other traffic participants including their

behavior or information from V2X communication.

Before any simulation can take place, the

simulations must first be configured. Thus, the

control of the simulation environment is divided into

the configuration of the simulation cycles on the one

Control of the Simulation einvironment

Control and Monitoring of the Simulation

• Data Recording and Visualization

• Automated Simulation Execution

• Automatic Model Update

• ML Cycle Execution

• Termination Conditions

Configuration of the Simulation Cycles

• Scenario Generator

• ANN Generator

• ML Configuration

Execution Control

Simulation Model

Ego Vehicle

• Driving Function

• Sensor Models

• Vehicle Dynamics Models

• Component Models

• Communication

Scenario/Environment

• Road Model

• Participants

Joint Simulation

Workspace

Development of a Simulation Environment for Automated Model Conﬁguration for the Design and Validation of AI-based Driving Functions

167

hand, and the simulation control and monitoring on

the other hand. The configuration of the simulation

cycles is basically the input interface of the user. With

the scenario generator it is possible to select how

exactly each simulation should look and run

according to the function specification. For example,

the route, the type, number and movement of other

road users, etc. can be set. Of course, the composition

and modeling depth of the ego vehicle must also be

defined.

As mentioned, this simulation environment can

be used to design not only AI-based functions, but

also functions based on conventional methods.

However, since this paper focuses on the use of ANN,

the connection of the simulation to their training will

now be described. After the general setting of the

simulations has been determined, it is important to

first configure the ANN architectures to be used. The

ANN generator allows a convenient configuration of

these ANNs in order to insert them into the simulation

as a model component (driving function). The

training of the ANNs is always iterative and requires

the execution of several simulations in the whole

scenario in which the ANN is used with the respective

architecture and the associated weights. The goal of

the training or the simulations is the successive

improvement of the ANN performance with respect

to the development requirements. The simulations run

according to a specific scheme depending on the

training algorithm and training parameters. The

configuration of the training in the ML Configuration

allows the simulations (also ML cycles) to be

automatically adapted and executed according to this

scheme. In summary, the information from the user

input is used as execution control for the “Control and

Monitoring of the Simulation” (Figure 1).

Thus, an internal flowchart is generated to

automate the simulation cycles. In it, for example,

several ML cycles are started first in one scenario and

then in another. Before each cycle, the corresponding

parameters of the active configuration are loaded and

transferred to the simulation model via the MATLAB

workspace. In this way, the model is automatically

updated. The workspace serves as an interface

between the simulation control and the simulation

model. It enables the recording and storage of

simulation data for visualization and later evaluation.

Furthermore, it transmits selected variables for model

monitoring. In case of previously defined,

inadmissible conditions, the simulation is

automatically aborted.

6 SIMULATION AND

EVALUATION

6.1 Description of the Use Case and

Modeling

To demonstrate this, a driving function for automated

lateral guidance at constant speeds on arbitrary, one-

lane routes without other road users is to be designed

in this use case. The example is consciously chosen

in a compressed way to show the simulation and time

effort. This is to illustrate the benefit of the automated

model configuration.

Since the ego vehicle’s speed is constant in this

application, a linear single-track model is used to

represent the vehicle dynamics. Automatically

generated routes according to the guidelines of the

German Federal Highway Research Institute with a

constant width of 3.5 m form the environment. The

sensor model consists of eleven lines, which detect

the distance to the lane boundaries in an angular range

of ±40 ° and a radius of 8 m. Figure 2 a) and b)

illustrate the use case with the visualization function

of the simulation environment. Other vehicle

components and communication systems of the ego

vehicle are not considered. The automated lateral

guidance is executed by an ANN in the simulation.

The eleven sensor values are the inputs, a

corresponding steering angle is the output.

The ANN is supposed to learn a natural steering

behavior with a self-implemented RL method, so

called Genetic Algorithms (GA). Natural steering

behavior in this case means constantly keeping the

center of the road and avoiding strong or high-

frequency oscillations to ensure safe and comfortable

driving behavior. GAs are a group of algorithms that

imitate the natural process of evolution in order

to successively approach an optimal solution. A so

Figure 2: Overview of the use case and modeling.

SIMULTECH 2021 - 11th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

168

called population, which consists of several solution

candidates (individuals), evolves generation by

generation by selection, recombination and mutation.

A population size of 50 means here that 50 different

ANNs are simulated in one generation. The iterative

character of the GA tends to result in relatively many

simulations.

Figure 2 c) shows an overview of the basic

structure of the simulation model. In a simulation, the

ego vehicle is placed on the track or scenario. There,

in one simulation step, the sensor model implicitly

captures the position and orientation of the ego

vehicle on the track via the eleven sensor lines. The

eleven sensor signals are passed to the ANN which

calculates a steering angle. This is used in the vehicle

dynamics model to calculate a new vehicle position

on the track. This process is repeated until the ego

vehicle reaches the destination or exceeds one of the

road markings. The GA is actually located outside the

simulation model. The associated reward function

accesses simulation data to evaluate the behavior of

individuals. The GA uses this information to update

the connection weights of the ANN. For this purpose,

several simulation runs are performed according to a

certain scheme of the ML cycle (Section 6.2).

6.2 Configuration of the Simulation

Cycles

As described in section 3, the determination of the

ANN architecture and the parameters of the ML

algorithm is always empirical. Furthermore, it is

important to pay attention to the generalization

capability when training ANNs. In this case, this

means that the ANN must be able to perform the

automated lateral guidance even on unknown routes.

To consider this aspect directly during the training,

several test runs are performed directly after every

single training. Thus, it is necessary to configure

several different simulations.

In order to successively achieve an ideally

designed driving function, multi-step tests must be

implemented in the configuration of the simulation

cycles. The simulation cycle always consists of

several ML cycles. One ML cycle always includes a

training of one ANN architecture with one GA

parameter set on one route with one reward function,

followed by four test runs on additional routes. While

the training phase of the GA consists of many

separate simulations for the individuals and

generations, only the best individual from the training

is tested in each of the test runs. Thus, the number of

simulations 𝑆



to be performed per ML cycle 𝑖 is a

function of the population size 𝑃



, the number of

Figure 3: Sequence of the automatic simulation execution.

generations 𝐺



and number of test runs 𝑇



. The

number of simulations in the total cycle 𝑁



is the sum

of all 𝑆



for the different ML cycles or configurations

𝐾



𝑁



=𝑆



∈



=𝑃



∙𝐺



+𝑇

∈



(1)

Assuming a population size of 50 and a

generation number of 25, 1254 simulations are

consequently performed in one ML cycle.

The first configuration involves the investigation

of six variants of the GA parameter sets for training a

base ANN (Figure 3 a)). The base ANN is a non-

optimized ANN that is assumed to be able to perform

the function fundamentally. The same is true for the

basic reward function. The resulting optimal GA

parameter set is used to determine the best of twelve

preconfigured ANN architectures, as shown in Figure

3 (b). In the final step (Figure 3 c)), the actual

optimization of the ANN behavior is performed by

performing further ML cycles with nine different

reward functions. According to equation (1), the total

number of simulations 𝑁



performed is in the high

six-digit range. Without the automatic model

configuration of the automated simulation

environment, the required effort would have

exceeded a reasonable level.

a) ML Cycles

Parameters Variants

ANN-Architecture* 1

Route 1…5

Reward Function 1

GA-Parameters 1…6

b) ML Cycles

Parameters Variants

ANN-Architecture* 1…12

Route 1…5

Reward Function 1

GA-Parameters best (a)

c) ML Cycles

Parameters Variants

ANN-Architecture* best (b)

Route 1…5

Reward Function 1…9

GA-Parameters best (a)

*Parameter to be Optimized

Development of a Simulation Environment for Automated Model Conﬁguration for the Design and Validation of AI-based Driving Functions

169

Figure 4: Simulation results of the use case on a test track.

6.3 Simulation Results and Evaluation

After the automatic execution of all configurations

from Figure 3, a function for automated lateral

guidance has been created. The ANN can safely and

comfortably take over lateral guidance on any route

in a speed range from 30 to 70 km/h. Corresponding

simulation results are shown in Figure 4. There, the

lateral deviation from the middle of the lane Δ𝑦

(Figure 4 a)), the steering angle 𝛿 (Figure 4 b)) and

the steering angle velocity 𝛿



(Figure 4 c)) are shown

for different fitness functions over the x-coordinate of

a test track (Figure 4 d)). The figure thus illustrates a

subsection of the ML cycle from Figure 3 c) for

optimizing the behavior with the reward function.

The best driving behavior in terms of comfort and

safety was achieved with the 𝐹𝑖𝑡

%

function

(yellow) due to the lowest lateral deviation and

oscillations. In this fitness function, the ego-vehicle

or ANN was rewarded for moving forward on the

track as well as for reaching the destination. It was

penalized for deviations from the middle of the lane

as well as for large steering angle changes. For more

detailed descriptions of the design and validation of

the driving function from this use case, please refer to

the previous work (Yarom et. al., 2020a) and (Yarom,

Jacobitz and Liu-Henke, 2020b).

The focus of this paper is on the simulation

environment. For its evaluation, a single automated

ML cycle was hand-programmed in MATLAB and

compared with the simulation environment in terms

of runtime. In the hand-programmed version, all

Figure 5: Comparison of simulation times in the use case.

model components according to Figure 2 c) as well as

the visualization were implemented. The comparison

according to Figure 5 showed that a single ML cycle

in the hand-programmed environment takes about 22

h and 36 min. In the simulation environment for the

automated model configuration, on the other hand, an

ML cycle with identical configuration takes about 28

min, which is only about 2% of this time.

Furthermore, it is possible to switch off the

visualization in the simulation environment. In this

way, the time required for each ML cycle can be

further reduced to a quarter.

The simulation environment presented in this

paper now makes it possible, on the one hand, to

design automated vehicle functions based on ANN.

On the other hand, all further requirements from

section 4.2, e.g. optimized computing time,

compatibility, usability and automation, have been

implemented. Thus, it exactly fulfills the originally

intended purpose of uniting the worlds of "modeling

and design of automated vehicles" and "AI

development" or their respective advantages. Thus, a

tool has been created with which the development and

validation of automated vehicles can be developed

safely and efficiently according to the methodology

presented in section 2. The result is a significant

contribution to the progress of the project and

autonomous driving.

7 CONCLUSION AND OUTLOOK

In this paper, a simulation environment for automated

model configuration for the design and validation of

AI-based driving functions was presented. Starting

with an introduction and motivation, the design

methodology and the state of the art were presented.

From this, the necessity for the development of the

presented simulation environment was derived and

requirements for it were defined. The implementation

of the requirements was described in the concept and

the realization. The result is a simulation environment

based on MATLAB and Simulink, which has a high

degree of compatibility with existing development

environments. In addition, it can automatically

0 100 200 300 400 500

-0.4

-0.2

0.2

0.4

a) y in m

Fit

100%

Fit

50%

Fit

80%

max

0 100 200 300 400 500

-4

-2

b) in °

0 100 200 300 400 500

x-coordinate of the test track in m

-400

-200

200

400

c) steering angle speed (steering wheel) in ° s

-1

d) Test track

0 500 1000 1500

Simulation Environment

(No Visualisation)

Simulation Environment

Hand-Programmed

Simulation Time in min

≈ 22 h 36 min

≈ 28 min

≈ 7 min

≈ 2 %

≈ 0,5 %

SIMULTECH 2021 - 11th International Conference on Simulation and Modeling Methodologies, Technologies and Applications

170

execute and visualize large simulation series in a

short time after a user-friendly configuration. This

makes it ideally suited for use in the design

methodology presented. Finally, the simulation

environment was used in an example application,

demonstrating its benefits and functionality.

Individual results of the application as well as their

relevance for the simulation environment were

presented and critically discussed. Future work steps

include extending the model and function library and

integrating it with dSPACE ASM.

ACKNOWLEDGEMENTS

This publication resulted from the subproject

"autoEMV" (Holistic Electronic Vehicle

Management for Autonomous Electric Vehicles) in

the context of the research project "autoMoVe"

(Dynamically Configurable Vehicle Concepts for a

Use-specific Autonomous Driving) funded by the

European Fund for Regional Development (EFRE |

ZW 6-85030889) and managed by the project

management agency Nbank.

REFERENCES

Alaoui, C. (2019). Hybrid Vehicle Energy Management

Using Deep Learning. 2019 International Conference

on Intelligent Systems and Advanced Computing

Sciences (ISACS), Taza, Morocco.

Deter, D., Wang, C., Cook, A., Perry N. K. (2021).

Simulating the Autonomous Future: A Look at Virtual

Vehicle Environments and How to Validate Simulation

Using Public Data Sets. In IEEE Signal Processing

Magazine, vol. 38, no. 1.

Duriez, T., Brunton, S., Noack, B. R. (2017). Machine

Learning Control. Springer International Publishing,

Cham, Switzerland.

Fayjie, A. R., Hossain, S., Oualid D., Lee, D. (2018).

Driverless Car: Autonomous Driving Using Deep

Reinforcement Learning in Urban Environment. 2018

15th International Conference on Ubiquitous Robots

(UR), Honolulu, Hawaii.

Huang, Z., Xu, X., He, H., Tan, J., Sun, Z. (2019).

Parameterized batch reinforcement learning for

longitudinal control of autonomous land vehicles. In

IEEE Trans. Syst., Man, Cybern, Syst., vol. 49, no. 4.

Kukkala, V. K., Tunnell, J., Pasricha, S., Bradley, T.

(2018). Advanced Driver-Assistance Systems: A Path

Toward Autonomous Vehicles. In IEEE Consumer

Electronics Magazine. vol. 7, no. 5.

Kuutti, S., Bowden, R., Jin, Y., Barber, P., Fallah, S.

(2021). A Survey of Deep Learning Applications to

Autonomous Vehicle Control, In IEEE Transactions on

Intelligent Transportation Systems, vol. 22, no. 2.

Milz, S., Schrepfer, J. (2020). Is artificial intelligence the

solution to all our problems? Exploring the applications

of AI for automated driving. In Bertram T. (eds)

Automatisiertes Fahren 2019. Springer Vieweg,

Wiesbaden, Germany.

Liu-Henke, X., Scherler, S., Fritsch, M., Quantmeyer, F.

(2016). Holistic development of a full active electric

vehicle by means of a model-based systems

engineering. 2016 IEEE International Symposium on

Systems Engineering (ISSE), Edinburgh, UK.

Lyu, H., Fu, H., Hu, X., Liu, L. (2019). Edge-Based

Segmentation Network for Real-Time Semantic

Segmentation in Traffic Scenes. 2019 IEEE

International Conference on Image Processing (ICIP),

Taipei, Taiwan.

Skansi, S. (2018). Introduction to Deep Learning.

Undergraduate Topics in Computer Science. Springer,

Cham, Switzerland.

Stančin I., Jović A. (2019). An overview and comparison of

free Python libraries for data mining and big data

analysis. 2019 42nd International Convention on

Information and Communication Technology,

Electronics and Microelectronics (MIPRO), Opatija,

Croatia.

Tirumala, S.S. (2020). Evolving deep neural networks

using co-evolutionary algorithms with multi-population

strategy. In Neural Comput & Applic, vol. 32.

Togelius, J., Juul, J., Long, G., Uricchio, W., Consalvo, M.

(2018) What Is (Artificial) Intelligence?. In Playing

Smart: On Games, Intelligence, and Artificial

Intelligence. MIT Press.

Wang, D., Devin, C., Cai, Q. -Z., Yu, F., Darrell, T. (2019).

Deep object centric policies for autonomous driving.

2019 International Conference on Robotics and

Automation (ICRA), Montreal, Canada.

Yarom, O. A., Scherler, S., Goellner, M., Liu-Henke, X.

(2020a). Artificial Neural Networks and Reinforcement

Learning for model-based design of an automated

vehicle guidance system. 12th International

Conference on Agents and Artificial Intelligence

(ICAART), Valletta, Malta.

Yarom O. A., Jacobitz S., Liu-Henke X. (2020b). Design of

Genetic Algorithms for the Simulation-Based Training

of Artificial Neural Networks in the Context of

Automated Vehicle Guidance. 2020 19th International

Conference on Mechatronics - Mechatronika (ME).

Prague, Czech Republic.

Zhang, J., Cho, K. (2017). Query-efficient imitation

learning for end-to-end autonomous driving.

Proceedings of the Thirty-First AAAI Conference on

Artificial Intelligence (AAAI-17), San Francisco, USA.

Development of a Simulation Environment for Automated Model Conﬁguration for the Design and Validation of AI-based Driving Functions

171