Systematic Model-based Design of a Reinforcement Learning-based

Neural Adaptive Cruise Control System

Or Aviv Yarom

, Jannis Fritz

, Florian Lange and Xiaobo Liu-Henke

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

Keywords: Design Methodology, Artificial Neural Networks, Automated Driving, ACC, Automated Longitudinal

Guidance, Artificial Intelligence, Neural Architecture Search, NAS.

Abstract: In this paper, the systematic model-based design of a reinforcement learning-based neuronal adaptive cruise

control is described. Starting with an introduction and a summary of current fundamentals, design methods

for intelligent driving functions are presented. The focus is on the first-time presentation of a novel design

methodology for artificial neural networks in control engineering. This methodology is then applied and fully

validated using the example of an adaptive cruise control system.

1 INTRODUCTION

The innovation alliance autoMoVe (Dynamically

Configurable Vehicle Concepts for Use-Specific

Autonomous Driving) (Raulf et. al., 2020), funded by

the European Regional Development Fund (ERDF)

(European Commission, 2014), aims to develop an

autonomous, modular and electric vehicle concept.

By exchanging application-specific modules at

runtime, a wide range of applications from internal

freight transport to passenger transport in road traffic

is to be realized autonomously. Within the framework

of this research project, the Ostfalia autoEVM sub-

project (Holistic Electronic Vehicle Management for

Autonomous Electric Vehicles) focuses on the

model-based development of innovative intelligent

algorithms and functions for autonomous driving.

With a higher degree of automation in driving, the

requirements for the vehicle or the automated driving

functions also increase. Current functions and

algorithms based on methods of control theory or

classical information processing can no longer fully

meet these requirements (Milz and Schrepfer, 2020).

Artificial intelligence (AI) therefore represents a key

technology for many domains involved in the

development and usage of intelligent, automated

vehicles (Fayjie et. al., 2018).

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

https://orcid.org/0000-0001-5613-4129

Modern vehicles and driving functions are,

regardless of the type of information processing,

complex mechatronic systems with a high degree of

internal and external networking. To master this

complexity in the development and validation

process, a systematic design approach is essential. On

one hand, a verification-oriented and simulation-

based design methodology established in

mechatronics research is used. (Liu-Henke et. al.,

2021) AI algorithms, especially artificial neural

networks (ANN) used in this work, are very different

from control theory approaches in their functionality

and design process. Therefore on the other hand, a

novel design methodology for ANNs is applied and

presented for the first time in this paper.

As a sub-goal of the above mentioned research

project, the present work applies the mentioned

design methodologies for ANNs and mechatronic

systems on a function for automated longitudinal

guidance in terms of intelligent adaptive cruise

control. The term intelligent refers on the one hand to

the use of ANNs, which are modeled according to the

functioning of the human brain, and on the other hand

to the self-learning or experience-based training

approach of reinforcement learning.

Yarom, O., Fritz, J., Lange, F. and Liu-Henke, X.

Systematic Model-based Design of a Reinforcement Learning-based Neural Adaptive Cruise Control System.

DOI: 10.5220/0010923300003116

In Proceedings of the 14th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2022) - Volume 3, pages 889-896

ISBN: 978-989-758-547-0; ISSN: 2184-433X

889

2 STATE OF THE ART

2.1 Intelligent Driving Functions for

Automated Longitudinal Guidance

With an increasing number and interconnection of

advanced driver assistance systems (ADAS), the

human driver successively delegates driving tasks to

the vehicle until he finally becomes a passenger in

autonomous driving. At that point, one no longer

speaks of ADAS, but of (automated) driving

functions. One example of such an automated driving

function is adaptive cruise control (ACC). In contrast

to simple cruise control, in which only a comparison

between the set speed of the driver and the actual

speed of the vehicle takes place, ACC can overwrite

the set speed downwards by evaluating

environmental sensors. The vehicle then maintains a

speed-dependent safe distance to vehicles and objects

in front by specifying reference values to subordinate

systems of the longitudinal dynamics (drive, brake,

transmission). (Lin, Nguyen and Wang, 2017)

Radar sensors are typically used for detecting

obstacles ahead and the subsequent determination of

relative speed and distance (Abdullahi and Akkaya,

2020). Modern systems often additionally use camera

and lidar sensors for object detection. Another

approach is the use of wireless vehicle-to-everything

(V2X) communication to realize cooperative driving

operation (Anayor, Gao and Odekunle, 2018).

From a control engineering point of view, ACC is

a cascaded system in which the inner cascade controls

a speed given by the outer cascade to regulate the

safety distance. In practice, however, some driving

functions exist that switch between distance keeping

and pure speed control. But for highly automated

driving, higher-order and holistic functions are

advantageous. (Abdullahi and Akkaya, 2020) In the

current literature, many approaches, such as a

classical PI (Kiencke and Nielsen, 2005), non-linear

model predictive (Shakouri and Ordys, 2014), or even

neuro-fuzzy control (Lin, Nguyen and Wang, 2017)

can be found for an ACC, each with specific

advantages and disadvantages. In this paper, a novel

holistic approach for an ACC using ANNs and

reinforcement leaning will be investigated.

2.2 Basics of Artificial Neural

Networks and Machine Learning

The term AI covers a large number of different

methods and algorithms that deal with the

autonomous and automated solving of problems

(Fayjie et. al., 2018). ANNs and ML form a subfield

of AI that has proven to be suitable for 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 creates the ANN.

Combinations of up to several hundred 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. The optimal

architecture of an ANN cannot be determined

analytically so far (Tirumala, 2020). Therefore,

besides experience and test series, a systematic design

methodology is necessary to find a suitable

architecture in the trade-off between computational

effort and performance capability. The number,

interconnection and weighting of the 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 presented with input data and the

associated output. The ANN iteratively learns the

relationship between these two quantities (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 method is used when no training data is

available, e.g. in automated vehicle guidance (Huang

et. al., 2019). SL and RL are upper categories of

learning methods, with diverse training algorithms.

Just like the ANN architecture, the optimal training

algorithms as well as their parameters cannot be

determined analytically. Thus, experience and

experimentation are required here as well.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

890

3 DESING METHODOLOGY

3.1 Model-based Controller Design

The complexity of modern vehicles is constantly

increasing due to the higher degree of internal and

external interconnection and the growing number of

intelligent and powerful hardware and software

components. In order to master the system

complexity and to avoid errors at an early stage in the

design of information processing, a holistic design

methodology is indispensable. Therefore, the holistic,

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 become established. (Liu-Henke et.

al., 2016)

The methodology builds on function-oriented

physical models of a controlled system. The control

function is then simulatively designed 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 (Jacobitz and

Liu-Henke, 2020). The subsequently automatically

generated function code is tested again against the

plant 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 (Liu-Henke et. al., 2016). 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 high

resource requirements 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., 2020). They enable feasible

and reproducible tests at any time without direct

dependency on physical prototypes, day times or

human experts. Thus, simulation runs can be

automated for different functional variants or

scenarios. This makes this methodology particularly

suitable for training ANNs. This is because, with a

few exceptions, machine learning is always iterative.

Virtual design methods like this form the basis for

many intelligent systems, such as highly automated

vehicles. With prototype-based testing, the hundreds

of thousands of test miles required cannot be

accomplished with reasonable time and cost. It should

be noted that despite the use of this methodology,

prototype-based tests cannot be completely

eliminated, for example for certification tests.

However, the number and the effort as well as the

associated disadvantages can be reduced to a

minimum. (Yarom et. al., 2020)

3.2 Systematic Design of ANNs

The design methodology for model-based controller

design presented in the previous section 2.1 is a

superordinate methodology. The design of an ANN

can be considered as a part of this methodology,

analogous to a controller design with dynamic

compensation. ANN design is a complex process

consisting of many design decisions, which

individually or in combination have a major impact

on the design outcome. These decisions are

distributed over the entire development process of the

ANN and concern e.g. the system structure (end-2-

end function / combination of sub-functions), the

learning method (SL / RL), the network architecture,

the parameterization of the training algorithm or the

evaluation of the ANN.

The special characteristic lies in the fact that in

contrast to the very systematic-analytical design

procedures of classical and modern control

engineering, the decisions and their effects in ANN

design are usually not mathematically-logically

understandable for the human mind. As a result, there

is no systematic design procedure for ANNs

(Tirumala, 2020). Although one can find many

applications of ANNs in current literature,

information on the respective design processes is rare.

When information is provided, the authors often

follow a result-oriented empirical approach based on

experience with low systematics.

The procedure model shown in Figure 1 presents

a first approach to systematize the design process of

ANNs. The starting point is the control problem

specification. It hardly differs from the classical

control theory. For example, especially in the case of

RL, the work is typically model-based. The special

feature is the choice of the system structure and

interfaces. While these are often determined by

physical conditions in classical controllers, they can

theoretically be freely selected for ANNs.

Subsequently, a suitable learning approach or a

concrete learning algorithm must be chosen for the

respective problem. There are many learning

algorithms, with specific advantages and

disadvantages, which have to be selected for each

application. In RL, a basic distinction is made

between gradient-based and gradient-free algorithms.

Systematic Model-based Design of a Reinforcement Learning-based Neural Adaptive Cruise Control System

891

Figure 1: Procedure model for the systematic design of

ANNs

The selection of the network architecture is the

most complex part of the ANN design and has

become a modern field of research in its own right. It

is attempted to formulate the architecture selection as

a search problem and to solve it automatically by

means of a higher-level optimization technique. This

is called Neural Architecture Search (NAS). (Rock et.

al., 2021) After the network architecture has been

determined, one can continue with the training, which

is the actual core of the design. In this process, the

behavior of the ANN is optimized with respect to the

original purpose. In this step, especially the

hyperparameters of the training algorithms have to be

set. The last step is the evaluation of the ANN with

respect to the original requirements and considering

its application purpose, e.g. on a real-time system.

As already indicated for the NAS, each block in

the process model from Figure 1 forms a higher-level

process step and each contains further complex

processes and decisions. Since this paper focuses on

the application of an ANN as an ACC, only some of

these will be taken up in more concrete terms later in

this paper.

4 SYSTEMATIC MODEL-BASED

DESIGN OF NEURAL

ADAPTIVE CRUISE CONTROL

The aim of this paper is the systematic model-based

design of a reinforcement learning-based neural

adaptive cruise control system. This process is now

presented according to the design methodology

presented in Figure 1.

4.1 Control Engineering Problem

Specification

4.1.1 Requirements Definition

The automated longitudinal guidance function

presented and developed here is intended to meet the

following requirements:

 Maintain a speed set by the driver and in no case

exceed it

 Maintain a 2-second safety distance to a vehicle

that may be driving ahead

 Do not fall below the safety distance, or only

slightly and for a short period of time

 Follow any speed profile, even when

maintaining the safety distance

 Maintain speed and distance without

oscillations if possible

•Training Data

•Collect

•Preprocess

•Reward Function

•Set up Function

•Determine Termination Conditions

•Parameterize the Algorithms

•Consideration of Constraints

Training

•Requirements Definition

•Interface Definition

•Modeling

•Analysis

•System Structure

Control Engineering

Problem Specification

•Determine and Check Evaluation

Criteria

•Computational Effort

•Real-Time Capability

•Control Deviation, Stability...

•Restrictions

•Generalization Capability / Robustness

Evaluation

•Network Type

•Number of Layers

•Interconnection of the Layers

•Layer Size

•Tapped Delay Lines

•Activation Functions

Network Architecture

•Supervised

•Unsupervised

•Reinforced

•Concrete Algorithm

Learning Mode and

Training Algorithm

Further complex processes

and decisions

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

892

 The ego vehicle should realize accelerations in

the range -3 m∙s-2 to 2 m/s2

 Be based on ANNs and RL

4.1.2 System Structure and Interface

Definition

Figure 2 schematically shows the system structure of

the neural adaptive cruise control designed here,

including the system environment. The starting point

is a lead car, which moves with an arbitrary velocity

lead

and thus changes its position x

lead

The actual

center of the structure, however, is the ego vehicle

model, which calculates the velocity v

ego

and position

ego

of the ego vehicle based on acceleration forces F

and braking forces F

. These positions and velocities

of the ego and lead car, are used by the radar sensor

with "built-in" preprocessing to determine, within its

range of 150 m, the input variables for the ANN.

These are:

 The velocity error verr, i.e. the difference

between the velocity vset specified by the driver

and the actual ego velocity

 The integral of the speed error over time ∫ verr

 The ego speed v

ego

Figure 2: Functional structure of the reinforcement

learning-based neural adaptive cruise control system.

 The distance error d

err

, i.e., the difference

between the safe distance d

safe

and the actual

distance between the vehicles d.

Using these input variables, the ANN then

calculates a command value for the driving and brake

pedal position p ∈ {-1 ≤ p ≤ 1 | ℝ}, which is

converted into the above-mentioned longitudinal

forces by a subordinate vehicle dynamics control

system.

These are all components that the system uses in

the application phase, i.e. after training. During

training, two additional components are needed. The

reward function, specified or to be worked out by the

developer, which calculates a reward based on data

of the system components in order to evaluate the

ANN during training. The RL algorithm is an

optimization technique that adjusts the weights w

i,j

the ANN to maximize the expected reward in the

upcoming episode (one simulation run).

4.1.3 Modeling

The overall system shown in Figure 2 is modeled,

simulated and trained in the simulation environment

for the automated model configuration for the design

and validation of AI-based driving functions from

(Yarom and Liu-Henke, 2021). A classical

longitudinal dynamics model is used to represent the

behavior of the Ego and Lead vehicles:

𝑚∙𝑥 𝐹



𝑐



∙

𝐴

∙

𝜌

∙𝑥













 .

𝑚∙sin



𝛼













 .

 𝑚 ∙ cos



𝛼



∙

𝑓













 .

𝐹



(1)

The variable for the driving and brake pedal

position p is scaled to the corresponding forces F

and

. In order to be able to simulate a standstill and

reversing, further mechanisms were implemented in

the simulation environment, which will not be

discussed in this paper.

The radar sensor with preprocessing is activated

as soon as the distance between the vehicles falls

below the range of the sensor. Here, on the one hand,

situation-dependent reference values for distance and

speed are calculated. For example, min(v

set

ego

)

applies to the desired speed of the ego car if the ego

car is in the range of d

safe

to the lead car or below. On

the other hand, the calculated set points are also

compared with the actual values and deviations are

calculated for the distance and the ego speed. Thus,

this function component not only provides the input

variables for the ANN, but also basic variables for the

reward function.

Systematic Model-based Design of a Reinforcement Learning-based Neural Adaptive Cruise Control System

893

4.1.4 Learning Mode and Training

Algorithm

In principle, only RL algorithms can be used for this

application. In order to create an optimal ACC

function on the one hand and to gain further

experience for the systematic design of ANNs on the

other hand, different training algorithms will be used

and compared. In preliminary tests with lower

requirements the following training algorithms were

used:

 Genetic Algorithm (GA)

 Particle Swarm Optimization (PSO)

 Deep Deterministic Policy Gradients (DDPG).

The first two methods are gradient-free,

evolution-based algorithms. This class of algorithms

usually evolves steadily toward a higher reward, but

does not always converge to a true optimum. Their

advantage is that you usually get to a good result

fairly quickly. A special feature of the GA is that it is

quite random and tends to find the global optimum

rather than other algorithms if the parameters are set

correctly. The DDPG, on the other hand, is a gradient-

based algorithm that converges fairly reliably to a

local optimum. However, it usually takes more time

to do so. Moreover, it rarely reaches the global

optimum.

The series of experiments showed that the DPPG

achieves the best results with high probability. To

make the results comparable, the same rudimentary

reward function was used for all algorithms. While

this series of experiments was already used to

optimize the hyperparameters of the algorithms, the

reward function was designed in detail at a later point.

4.2 Network Architecture

By choosing DDPG as training algorithm, two ANN

architectures have to be chosen. This is because an

actor and a critic network are required here. The actor

network is the ANN that performs the actual control

task. The critic network serves as a translator between

the reward function and the algorithm for updating

the connection weights. It is only needed during the

training phase.

The critic network can usually be chosen as a

simple feedforward network with rectifier functions

as activation. Therefore, such an ANN with five

layers was specified here as the critic network. The

input variables are its output in addition to those of

the actor network. The architecture of the actor

network was automatically optimized using a NAS

based on a GA. The result is a feed forward with four

layers. In each of the first three layers there are 48

hidden neurons as well as rectifier functions for

activation. In the last layer there is only one neuron

with a tanh activation.

4.3 Training

Now that the network architectures have been

determined, the reward function that determines the

final behavior of the ANN or Ego Car must be

designed. This consists of four terms:

 -0.1∙(v

err

)

: Penalizing speed errors to maintain

the target speed.

 -(a)2: Penalization of accelerations, so that

oscillations and abrupt velocity changes are

avoided

 -0.1∙(derr)2: Penalization of distance errors so

that the target distance is maintained.

 +verr for verr ≤ 0.25 m/s: Reward of small

speed errors so that the target speed is

maintained.

In an iterative process, the ego car now runs

through several episodes in which the actor network

is used again and again as a longitudinal dynamics

controller and receives a corresponding reward for its

behavior. Figure 3 shows the course of the rewards

over the episodes as well as the average reward. The

reward is very negative at the beginning of the

training, since the ANN does not yet behave well

without prior knowledge. The reward quickly

improves to the range of about -600 and then slowly

increases until the ANN reaches about -91 reward

points in the best episode at the end. In between there

are occasional dips, which indicate a wrong weight

change, but can be compensated by the algorithm.

The fact that the best reward is -91 does not mean that

the behavior of the ANN is bad. This is only caused

by the design of the reward function. The course of

the reward is therefore more crucial than its absolute

value. Therefore, the evaluation after the training is

essential.

Figure 3: Course of the reward during training.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

894

4.4 Evaluation

The evaluation is the final step in the design of the

ANN for neural adaptive cruise control. A special

aspect that must be considered in the evaluation is the

so-called generalization capability. This means that

the ANN should apply the learned knowledge to new,

previously unknown problems. In order to be able to

rate the generalization capability, the numerous

necessary test scenarios for evaluation must therefore

differ from the training situation.

During training, v

set

was at 100 km/h and the lead

car was driving ahead with an acceleration sine wave.

In the one exemplary test situation shown here, v

set

increased to 108 km/h and the vehicle dynamics

model of the Lead Car is replaced by a "hard" speed

profile to further complicate the situation for the

ANN. The speed profile of the Lead Car is shown in

red in the upper part of Figure 4, v

set

in blue. The

yellow curve shows the speed of the Ego Car. In the

lower part of the figure, the distances d

safe

and d are

plotted versus time. At the beginning, the Ego Car

was deliberately placed very close (d < d

safe

) behind

the Lead Car, so that it must first maintain its speed

to keep the safe distance. Subsequently, the Ego Car

follows the speed curve of the Lead Car very well and

stably until the Lead Car accelerates to 120 km/h, i.e.,

above v

set

, at about 24 s. As desired, the ego car does

not exceed v

set

and the distance increases. Then, the

Lead Car brakes abruptly and the Ego Car also brakes

hard to maintain the safety distance without violating

the acceleration requirement (section 4.1). Finally,

the lead car accelerates again. The Ego Car controls

set

in a stable manner.

Figure 4: Simulation result of the neural adaptive cruise

control system.

This means that the ANN has learned the adaptive

cruise control and satisfies all requirements as

expected. Further simulation results, which could not

be presented here, have shown that the designed

function is generalization capable. Thus, the novel

methodology for systematic design of ANNs (in the

first phase) can also be considered validated.

5 CONCLUSION AND OUTLOOK

In this paper, the systematic model-based design of a

reinforcement learning-based neural adaptive cruise

control system was presented. Starting with an

introduction, the state of the art and the underlying

design methodology were summarized. This relates to

model-based controller design and systematic design

of ANNs. The latter was presented as a novelty for

the first time in this paper.

Subsequently, the application and validation of

this methodology was carried out on the example of a

neural adaptive cruise control system. The design

process was described in detail and the resulting

function was intensively evaluated and validated

against the requirements.

Future work steps include the further validation of

the design methodology as well as the design of

further automated driving functions.

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

Abdullahi, A., Akkaya, S. (2020). Adaptive Cruise Control:

A Model Reference Adaptive Control Approach. 2020

0 102030405060

100

110

120

130

v in km/h

set

lead

ego

0 102030405060

t in s

100

120

d in m

safe

Systematic Model-based Design of a Reinforcement Learning-based Neural Adaptive Cruise Control System

895

24th International Conference on System Theory,

Control and Computing (ICSTCC), Sinaia, Romania.

Anayor, C., Gao, W., Odekunle, A. (2018). Cooperative

Adaptive Cruise Control of A Mixture of Human-

driven and Autonomous Vehicles. SoutheastCon 2018,

St. Petersburg, FL, USA.

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

Learning Control. Springer International Publishing,

Cham, Switzerland.

European Commission (2014). European Regional

Development Fund. Available at: https://ec.europa.eu/

regional_policy/en/funding/erdf/ (Accessed: 22

December 2021).

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.

Jacobitz, S., Liu-Henke, X. (2020). The seamless low-cost

development platform LoRra for model based systems

engineering, 8th International Conference on Model-

Driven Engineering and Software Development

(MODELSWARD), Valletta, Malta.

Kiencke, U., Nielsen, L. (2005). Automotive control

systems: for engine driveline and vehicle. Springer,

Berlin, 2005.

Lin, Y.-C., Nguyen, H.-L. T., Wang, C.-H. (2017).

Adaptive neuro-fuzzy predictive control for design of

adaptive cruise control system. 2017 IEEE 14th

International Conference on Networking, Sensing and

Control (ICNSC), Calabira, Italy.

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.

Liu-Henke, X., Jacobitz, S., Scherler, S., Göllner, M.,

Yarom, O., Zhang, J. (2021). A Holistic Methodology

for Model-Based Design of Mechatronic Systems in

Digitized and Connected System Environments. 16th

International Conference on Software Technologies

(ICSOFT), online.

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.

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.

Raulf, C., Pethe, C., Vietor, T., Henze, R. (2020).

Dynamically Configurable Vehicle Concepts for

Autonomous Driving. ATZ Worldwide, vol. 122.

Rock, J., Roth, W., Toth, M., Meissner, P., Pernkopf, F.

(2021). Resource-Efficient Deep Neural Networks for

Automotive Radar Interference Mitigation. in IEEE

Journal of Selected Topics in Signal Processing, vol.

15, no. 4.

Shakouri, P., Ordys, A. (2014). Nonlinear model predictive

control approach in design of adaptive cruise control

with auto-mated switching to cruise control. Control

Engineering Practice, vol. 26.

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

using co-evolutionary algorithms with multi-population

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

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

(2020). 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., Liu-Henke, X. (2021). Development of a

Simulation Environment for Automated Model

Configuration for the Design and Validation of AI-

Based Driving Functions. 11th International

Conference on Simulation and Modeling

Methodologies, Technologies and Applications

(SIMULTECH), online.

ICAART 2022 - 14th International Conference on Agents and Artiﬁcial Intelligence

896