Conﬂicting Moral Codes for Self-Driving Cars: The Single Car Case

Ahmed Ibrahim , Yasmin Mansy, Nourhan Ehab and Amr Elmougy

Computer Science and Engineering Department, German University in Cairo, Cairo, Egypt

Keywords:

Conﬂict Resolution, Thomas Kilmann Model, Self-Driving Cars.

Abstract:

There has been incremental advances in the development of self-driving cars. However, there are still several

gaps that must be ﬁlled before a fully autonomous vehicle can be achieved. The ability to resolve conﬂicts

in the event of an unavoidable accident is one of the most prominent and crucial aspects of a self-driving car

that is currently absent. To address this gap, this paper aims to resolve moral conﬂicts in self-driving cars in

the case of an unavoidable accidents. Assuming we have a predeﬁned rule set that speciﬁes how a car should

morally react, any clash between these rules could result in a critical conﬂict. In this paper, we propose a novel

procedure to resolve such conﬂicts by combining the Thomas Kilmann conﬂict resolution model together with

decision trees. Evaluation results showcase that our proposed procedure excels in distinct ways, enabling the

self-driving car to make a decision that will yield the best moral outcome in conﬂicting scenarios.

1 INTRODUCTION

Driving a vehicle with minimum human intervention

is no longer science ﬁction (Cowger Jr, 2018). Cur-

rently, the automotive industry is undergoing a quan-

tum shift towards a future in which the driver’s role in

operating his or her vehicle will diminish to the point

of being eliminated (Barabas et al., 2017). For a self-

driving car to reach a stage where it would be fully

autonomous, it must be able to make faster decisions

under a wide range of situations, which could include

moral dilemmas such as choosing which potential vic-

tims to avoid. Such situations include unavoidable

accidents where life and death decisions have to be

made.

Although autonomous vehicles will eliminate the

major cause of human error, unavoidable deadly ac-

cidents, where the vehicle must make a life-or-death

decision in a fraction of a second, cannot be prevented

(Borenstein et al., 2017). In situations where moral

decisions need to be made by the vehicle, an ethical

framework must be implemented, where a predeﬁned

set of rules would allow the car to react. However,

if these rules contradict, the car would be unable to

determine a proper action, leading to a conﬂict. As

discussed later in this study, this must be resolved to

prevent catastrophic losses from indecision.

Accordingly, the objective of this research which

is to apply different conﬂict resolution techniques to

resolve conﬂicts that may arise due to various factors

that would be discussed in further details later in this

study, allowing the self-driving car to react in a timely

manner. As such, the ability to resolve conﬂicts that

may arise within a single self-driving vehicle has not

been previously addressed. In addition, the Thomas

Kilmann Model has primarily been used to resolve

conﬂicts between humans. Since we are attempting

to imitate a human-like trait, ethics, it stands to rea-

son that in order to imitate the ways in which con-

ﬂicts in ethics should be resolved, we would need to

employ another negotiation method used by humans,

which has not been attempted before to resolve con-

ﬂicts within self-driving vehicles using the Thomas

Kilmann Model in previous literature. Thus, the aim

is to provide a fast and accurate conﬂict resolution

technique that would allow self-driving cars to resolve

the conﬂict and react on time in case of an unavoid-

able accident. With the help of the Carla simulator

and ScenarioRunner, it is possible to build complex

conﬂicting scenarios where different resolution tech-

niques can be applied to imitate a real-world scenario

as closely as possible.

The rest of this paper is structured as follows. Sec-

tion 2 will present necessary background. Section

3 describes the proposed conﬂict resolution model.

Section 4 presents the evaluation of the proposed

model. Finally, Section 5 outlines some concluding

remarks.

Ibrahim, A., Mansy, Y., Ehab, N. and Elmougy, A.

Conﬂicting Moral Codes for Self-Driving Cars: The Single Car Case.

DOI: 10.5220/0011677500003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 1, pages 249-256

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

249

2 BACKGROUND

Unavoidable accidents are described as crucial situ-

ations in which no solutions can be developed in the

time available (to a human or self-driving car) to com-

pletely avoid the accident. The focus of this work is

on unavoidable accidents that result in fatalities, or

catastrophes in which human lives are lost. A self-

driving car must decide what action to take in the

event of an unavoidable accident, which includes de-

ciding on potential casualties. This is only partially

today investigated legally and ethically (Karnouskos,

2018).

2.1 The Need for Ethics

In a research paper by (Lin, 2016), the author provides

a basic scenario demonstrating the need for ethics in

self driving vehicles. When a self-driving car is faced

with a terrible choice in the distant future: where

it would hit an eight-year-old girl if it swerved to

the left, or an 80-year-old grandma if it swerved to

the right. Given the speed of the car, both victims

would almost probably die upon impact. If you do not

swerve, you will hit and kill both victims. Therefore,

there are sufﬁcient reasons to believe that you should

swerve in one of two directions (Lin, 2016). But how

should the vehicle react in such a scenario? There

may be reasons to pick one over the other, regardless

of how unappealing or distressing those reasons may

be. This is a dilemma that is not easily solvable, high-

lighting the importance of ethics in the development

of self-driving cars.

2.2 Conﬂicts in Decisions

Referring to the same case given in a research paper

by (Patrick, 2016), some may claim that hitting the

grandma is the lesser of two evils, at least in some

people’s eyes. As the girl has her entire life ahead of

her, whereas the grandmother has already lived a full

life and had her fair share of experiences, there are

reasons that seem to weigh in favor of saving the little

girl over the grandmother, if an accident is unavoid-

able.

But what if there were two girls of the same age

on each side of the road? Even if an autonomous car

with an ethical framework can select between the two

evils. Both sides would have equal priority according

to the car’s framework, having the same age, gender,

and any other aspect the car could compare to. Re-

sulting in a conﬂict since the car can no longer make

a decision. So even if we develop an ethical frame-

work capable of resolving such complex scenarios in

an attempt to replicate a human trait, conﬂicts will

still arise. Conﬂicts can arise for a number of reasons,

some of which are predictable and could be avoided,

while others we would not even know about.

Some might argue that future cars will not have to

make difﬁcult ethical decisions, that stopping the car

or handing control to the human driver is the easy way

around ethics. However, braking and handing over

control will not always be enough. These solutions

may be the best we have today, but if automated cars

are ever to be used more widely outside of limited

highway environments, they will need more response

capabilities (Lin, 2016).

In future autonomous cars, crash-avoidance fea-

tures alone will not be sufﬁcient. Sometimes an ac-

cident will be unavoidable as a matter of physics,

for myriad reasons–such as insufﬁcient time to press

the brakes, technological errors, misaligned sen-

sors, terrible weather conditions, and just pure bad

luck. Therefore, self-driving cars will require crash-

optimization strategies.

2.3 Thomas-Kilmann Conﬂict Model

In 1974, Kenneth W. Thomas and Ralph H. Kilmann

introduced the ThomasKilmann Conﬂict Mode In-

strument, which is presented as an alternative ap-

proach to the arising conﬂicts. The Thomas-Kilmann

model identiﬁes ﬁve distinct conﬂict resolution be-

haviours and how they effect problem resolution. As

illustrated in Figure 1, the various types of behaviour

reﬂect various degrees of assertiveness and coopera-

tion (Fahy et al., 2021).

Figure 1: Thomas-Kilmann Conﬂict Model.

The Thomas-Kilmann technique has primarily

been used to resolve conﬂicts in human-to-human ne-

gotiations. It has previously been used to resolve

conﬂicts between medical staff in pediatric surgery.

(Fahy et al., 2021). This is a real-time, critical,

and time-sensitive situation in which a decision made

could mean the difference between life and death.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

250

And must be made within a matter of seconds. This

is very similar to the situations we would face in the

event of an unavoidable accident, where the conﬂicts

must be resolved within very short deadlines. Fur-

thermore, since we are attempting to mimic a human-

like trait, ethics, it stands to reason that in order to

mimic the way conﬂicts in ethics should be resolved,

we would need to employ another human negotiation

method.

Next, we will analyse the ﬁve distinct types of be-

haviour and how they can be utilised to resolve an eth-

ical conﬂict in the event of an unavoidable accident.

1. Competing: Power-oriented, assertive, and unco-

operative. Having rules/decisions of equal impor-

tance. Each decision uses whatever power seems

appropriate to win that position, by competing on

any aspect favoring one side over the other.

2. Collaborating: assertive and cooperative. Usu-

ally involves two or more entities trying to ﬁnd a

solution that meets and satisﬁes both concerns.

3. Compromising: assertive and cooperative. Com-

promising splits the difference or exchanges con-

cessions to ﬁnd a quick middle-ground position.

4. Avoiding: unassertive and uncooperative. Know-

ing one of two conﬂicting decisions is wrong al-

lowing for a better decision to be considered.

5. Accommodating: is unassertive and coopera-

tivethe opposite of competingand involves an ele-

ment of self-sacriﬁce. During conﬂict resolution,

one side may realise its decision was wrong, thus

backing up allowing for a better solution.

3 CONFLICT RESOLUTION

Our proposed conﬂict resolution model consists of a

pipeline of three main stages: (1) Collecting diverse

data using a variety of sensors. Along with the for-

mulation of rules that would serve as the car’s ethical

foundation in the event of an unavoidable accident;

(2) Conﬂict resolution in case moral conﬂicts arise by

utilizing decision trees; and (3) Deciding which ac-

tions to take after the conﬂict resolution. The three

stages are described in this section.

3.1 Data & Rule Set

First we start by data collection, CARLA simulator

(Dosovitskiy et al., 2017) was used which is an open-

source autonomous driving simulator that utilize lot

of sensors that vehicles rely on to obtain data from

their surroundings, ranging from cameras to Radar,

LIDAR, and many more.

After we are done with data collected using dif-

ferent and various types of sensors, we then referred

to the MIT Moral Machine Experiment (Awad and

Dsouza, 2018) as our rule set. Allowing us to create a

18 priority based rule set as shown in Figure 2.

Having the top 9 rules indicated in blue as the

highest priority rules and the other nine indicated in

red as the lower priority rules. These priorities were

established in the following manner according to a re-

search paper by (Anbar, 1983) on estimating the dif-

ference between two probabilities.

(nd ∗ (1 − (highest∆P − current∆P)) + k

Where n is the number of the rules multiplied by a

factor d and ∆ P is the difference between the proba-

bility of sparing characters having the attribute in blue

and the probability of sparing characters having the

opposing attribute in red (for example Elderly/Young)

(Awad and Dsouza, 2018). And k is an offset that is

larger than the smallest expected negative number in

our rule set.

0 5 10 15 20 25 30 35

Action

Passengers

Males

Large

Lower status

Unlawful

Elderly

Pets

Fewer characters

Inaction

Pedestrians

Females

Fit

Higher status

Lawful

Young

Humans

More characters

Priority

Figure 2: Priority Based Rule Set.

3.2 Conﬂict Resolution via Decision

Trees

In this work, we use decision trees as means to re-

solve conﬂicts. A decision tree is an optimal choice

for data structuring since it allow for computational

efﬁciency and accuracy of classiﬁcation according to

a research paper by (Safavian et al, 1991) (Safavian

and Landgrebe, 1991).

3.2.1 Types of Conﬂicts

Decision trees classify any object, in our case a sce-

nario by moving it down the tree from the root to a

leaf node, with the leaf node providing the scenario’s

evaluation. Each node in the tree represents a test case

for some attribute, and each edge descending from

Conﬂicting Moral Codes for Self-Driving Cars: The Single Car Case

251

that node represents one of the test case’s possible an-

swers. This recursive approach is repeated for each

sub-tree rooted at the new nodes. A priority score can

be assigned at each level moving down the decision

tree and adds up until a leaf node is reached. The

situation with the highest ﬁnal priority score will be

given the top priority. There are two types of possible

conﬂicts that can occur:

1. Conﬂicting Nodes: this form of conﬂict would

occur within the decision tree nodes themselves.

We classiﬁed such types of conﬂicts into three

main categories.

(a) Comparable Branching: Where an attribute

depends on comparing its value to the other

lanes, such as more (versus less) characters,

having both lanes with the same number of

characters would cause a conﬂict and prevent

branching to either node.

(b) Mix of Characters: Having a mix of different

characters standing on the same lane, such as

having young and elderly ages standing on the

same lane in certain scenarios, prevents branch-

ing to either node, as illustrated in Figure 3.

the value for a certain attribute available at a

certain level, which could be due to various fac-

tors such as a lack of time or inaccurate sensors.

2. Conﬂicting Priority Scores: occurs when the de-

cision tree has been entirely traversed and the as-

signed scores for each lane are exactly equal.

Root

More Characters

Priority: 32

Priority: 1

Priority: 29

Priority: 4

Young Elderly

Fewer Characters

Priority: 29 Priority: 4

Young

Elderly

30 5

Figure 3: Conﬂicting decision tree.

3.2.2 Resolution Techniques

We use the Thomas Kilmann model techniques: com-

peting, avoiding, and accommodating. The remain-

ing two techniques (collaborating and compromising)

are not considered since they deal with multi-agent

cases, which is not the focus in this paper. Alongside

the above mentioned techniques, the verifying nodes

technique is used. Since the time required to traverse

the entire tree may exceed the time available before

the collision, the resolution strategies make use of a

deadline by which we must stop traversing the deci-

sion tree and assign the current scores, leaving extra

time to resolve any conﬂicts if both lanes have the

same priority scores. In what follows, we describe

the operation of the different resolution techniques on

decision trees.

Competing: implies contending over one or more as-

pects that might give one lane an edge over the other.

In our case the competition would be based on two

critical factors, time and highest priority score.

Root

Priority: 32 Priority: 1

More Characters

Priority: 29

Priority: 4

Young

Lawful

Elderly

Fewer Characters

Priority: 29

Priority: 4

Elderly

Young

Unlawful

Lawful

Unlawful

Lawful

Unlawful Lawful

Unlawful

Deadline

Figure 4: Deadline met before traversing entire tree.

Figure 6 illustrates a scenario in which we were

unable to traverse the entire decision tree as we hit the

deadline. Although we ended up with the same prior-

ity score, resulting in a conﬂict, it is uncertain whether

either of the lanes might have achieved a higher pri-

ority if we had enough time to complete the entire

decision tree.

As there is insufﬁcient time to traverse the entire

tree, the competing technique scans all of the leaf

nodes for each of the two lanes in search of the highest

possible priority score that could have been obtained.

And the one that have the highest possible score is

assigned a higher priority.

Competing is utilized in situations where time is

a prime concern, so time is considered to be a vital

factor. Each lane’s time to reach a ﬁnal node and

provide a score is recorded, regardless of whether

this ﬁnal node is a leaf or if the lane did not complete

the entire tree. The quicker a lane was able to receive

a score, the fewer conﬂicting nodes it encountered

while traversing the tree, resulting in it being faster

and having an advantage over the other lane by

completing the task in less time. And because an

unavoidable accident is an extremely time-sensitive

situation, it would be essential to have a lane with

fewer conﬂicts and a shorter computation time.

Avoiding: implies admitting that you were mistaken

or knowing that the other side has a better approach

to this circumstance (Schaubhut, 2007).

As shown in Figure 5, lane A contained all Pets,

while lane B contained humans, which should be pri-

oritised according to our rule set. However, lane A

received a higher Priority Score due to the fact that

the traversal of the decision tree was interrupted after

just two levels, resulting in both lanes having equal or

even higher priority to lane A.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

252

Root

Priority: 32 Priority: 1

More Characters

Priority: 31

Priority: 2

Human

Young

Pet

Fewer Characters

Priority: 31

Priority: 2

Pet

Human

Elderly

Young

Elderly

Deadline

Lawful

Figure 5: One of the two lanes is a leaf node.

In this case, the decision tree was not entirely

completed. The avoidance technique would be

applied by determining whether or not the ﬁnal node

for each lane has any children. If a lane came to a

stop at a leaf node. It should declare so and thus avoid

the conﬂict in the ﬁrst place, saving us more time

and resulting in a better outcome since the decision

would be made earlier.

Accommodating: is realizing that its own decision

was incorrect, allowing for a better solution to be con-

sidered (Schaubhut, 2007). It is also possible that

the current choice or node has been chosen most of

the time so it would accommodate to allow the other

nodes to contribute to the process of decision making

as well, given that both lanes have equal priority.

Root

More Characters

Young

Elderly

Fewer Characters

Elderly

Young

Correctness: 11/13

Frequency: 13

Correctness: 3/7

Frequency: 7

Correctness: 7/9

Frequency: 9

Correctness: 3/4

Frequency: 4

Correctness: 2/4

Frequency: 4

Correctness: 1/3

Frequency: 3

Figure 6: Correctness and frequency logs.

The main idea behind the accommodating tech-

nique is to record a history log, which is in charge of

two main things:

• Correctness Log. The correctness log stores in-

formation on comparable instances in the same

street or neighbourhood and the ratio at which

choosing this leaf or node was correct.

• Frequency Log. The Frequency log records the

number of times that leaf or node was chosen, al-

lowing for a balance in the frequency at which

each node is chose, given that both nodes have

equal priority.

In the event of a conﬂict if we have prior knowl-

edge about the accident location or the conﬂicting de-

cision tree nodes. The accommodating technique ex-

amines the node’s history for each processing lane, as

shown in 6. Starting with the Correctness ratio, if a

node’s correctness ratio is low, it allows for the other

lane to be selected. Similarly, if a node’s frequency is

higher, indicating that it has been selected more fre-

quently, it allows for the other lane to be selected.

Verifying Nodes: is a method inspired from (Bazan

et al., 2016) establishing a procedure for generating

trees with veriﬁcations of cuts deﬁned on numerous

attributes. The distribution of objects based on the op-

timal cut is conﬁrmed in each node of a tree formed

using this method by subsequent cuts on different at-

tributes that separate objects in a similar fashion.

Start

Only lane A is of a

mixed characters?

Characters on

lane B are of a higher

priority?

- Assign higher

priority to lane B.

- Assign lower

priority to lane A.

- Assign higher

priority to lane A.

- Assign lower

priority to lane B.

Ratio of lane A

is higher than ratio of

lane B?

- Assign higher

priority to lane A.

- Assign lower

priority to lane B.

Ratio of lane B

is higher than ratio of

lane A?

- Assign higher priority

to lane B.

- Assign lower priority to

lane A.

Assign zero to both lane

as both have the same

exact ratio, no lane have

an edge over the other.

Yes

Yes No

Figure 7: Verifying node ﬂowchart.

The verifying node technique resolve conﬂicting

nodes where we have a mix of characters standing

within the same lane preventing us from branching as

illustrated earlier in Figure 3.

As shown in Figure 7, we start by comparing one

of the two lanes, let’s call it lane A, to the other, lane

B. If only one of the two lanes contains a mix of char-

acters, we check to see if lane B contains characters

with a higher priority; if it does, lane B is assigned the

higher priority.

In the event that both lanes contained a mix of

characters, the lane with the higher ratio of high-

priority characters is assigned a higher priority.

For the two other conﬂicting nodes scenarios men-

tioned earlier: comparable branching and having at-

tributes with no value, as the values for both lanes

are equally identically or unknown, we skip that level

of the decision tree and assign zero score for both

lanes at that level, preventing it from affecting the ac-

curacy and ﬁnal priority scores of either lane.

3.3 Decision Making

The process of decision making depends on two main

factors:

• Time Remaining Before Collision. The time re-

maining determines whether we will be able to

Conﬂicting Moral Codes for Self-Driving Cars: The Single Car Case

253

traverse the whole tree, and whether we will be

able to run all the conﬂict resolution techniques

presented within the remaining time.

• The Complexity of the Scenario. Complexity of

the scenario, number of characters, lighting, and

other factors determine the time needed to process

this data and whether missing data will cause node

conﬂicts that must be resolved, taking more time.

Having one conﬂict resolution technique resolve

the issue is sufﬁcient to terminate and move forward

with this decision. Different resolution techniques

may be suitable in different situations, as some are

faster and more certain than others. In the next sec-

tion, we’ll examine all the different techniques and

determine which one excels in which aspects.

4 RESULTS & DISCUSSION

To ensure that the presented conﬂict resolution tech-

nique will be effective in a real-world scenario, and to

analyze and evaluate each of the techniques. This sec-

tion presents the method used to evaluate this project

and the results we obtained.

Carla simulation was used combined with Scenar-

ioRunner To simulate a real-world scenario as closely

as possible, generating conﬂicting scenarios in which

the proposed resolution techniques can be evaluated.

4.1 Evaluation Technique

Three different metrics were used to evaluate the efﬁ-

ciency of our conﬂict resolution techniques, that take

into account a range of relevant factors:

• Decidability. In this analysis, we adopt the de-

cidability index d (Williams, 1996), as a measure-

ment index of whether the technique is decidable

and would eventually result in a resolution.

• Correctness. Correctness is the accuracy of a de-

cision, and verifying a technique’s correctness can

be done by comparing its outcome to the top nine

priorities in Figure 2, and calculating a ratio of the

number of rules met against these top nine priori-

ties, which was adopted based on a related line of

research includes robust detection (Huber, 1965).

• Time. The time to reaction has been considered in

the literature (Wagner et al., 2018; Tamke et al.,

2011; Junietz et al., 2018) for which the time

needed by each technique to resolve a conﬂict

would be measured in a variety of complex sce-

narios, allowing us to determine whether a partic-

ular technique would be able to meet the deadline

and, if not, by how much it missed the deadline.

4.2 Testing Scenarios

As shown in Figure 8, all scenarios were ran three dif-

ferent times at three different distances to simulate a

shorter time remaining for reaction and conﬂict reso-

lution at each point. In each scenario, a random set of

characters were present on lanes A and B, where the

car was faced with a conﬂicting situation and must

react within the time available.

Figure 8: Evaluation Technique.

To ensure accurate results, a random set of 80 dif-

ferent conﬂicting scenarios were generated. Having

ﬁfty scenarios represent the case where the entire de-

cision tree could be traversed, while the remaining

thirty scenarios represent the case where there is no

time to traverse the entire decision tree before the

deadline. The results of the two cases are deeply ana-

lyzed in the following section.

4.3 Results

Figure 9 show an overall analysis after running 50

different scenarios that succeeded in traversing the

whole decision tree.

68%

94%

10%

0% 20% 40% 60% 80% 100%

Competing

Accommodating

Avoiding

None

Decidability

76%

64%

100%

0% 20% 40% 60% 80% 100%

Competing

Accommodating

Avoiding

Correctness

0.35s

0.1s

0.08s

0 0.2 0.4 0.6 0.8 1

Competing

Accommodating

Avoiding

Time

Deadline 15m

0.45s

Deadline 25m

0.75s

Figure 9: Results of traversing the whole tree.

Where each of the graphs represented in Figure 9

is discussed in more details below:

• Decidability. In 97% of the cases, we reached

a resolution, with the Accommodating technique

being the most successful, reaching a resolution in

94% of the scenarios, followed by the Competing

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

254

technique, with a resolution of 68%, and ﬁnally

the Avoiding technique, which only succeeded in

reaching a resolution in 10% of the cases.

• Correctness. Although the Accommodating

technique was the most effective in resolving most

scenarios, it was the least accurate, achieving only

64% accuracy, followed by the Competing tech-

nique with 76% accuracy, and ﬁnally the Avoid-

ing technique, which was the most accurate with

100% accuracy in an event of a conﬂict.

• Time. The Avoiding and Accommodating tech-

niques were the fastest in just under 0.1s, with the

Avoiding technique achieving faster averages by

0.08s. The competing technique was the slow-

est with 0.35s, yet it managed to ﬁnish before the

deadline of the shortest distance (0.45s).

4.3.1 Deadline met Analysis

Figure 10 show an overall analysis after running

30 different scenarios that were terminated before

traversing the whole decision tree.

Where each of the graphs represented in Figure 10

is discussed in more details below:

• Decidability. The results this time were rather in-

teresting, with a resolution being reached in 100%

of the cases. The Competing technique was the

most successful, resolving 96% of the scenarios,

followed by the Accommodating technique, with

a resolution of 92%, and ﬁnally the Avoiding tech-

nique, with a resolution of only 12%.

• Correctness. The Avoiding technique contin-

ues to be the most accurate with a highest score

of 100%, followed by the Competing technique,

which achieved 78% accuracy, and then the Ac-

commodating technique, which was once again

the least accurate with 67%.

• Time. The Avoiding and Accommodating tech-

niques still achieved nearly identical averages,

with the Avoiding technique remaining faster and

averaging 0.08s, while the competing technique

became even slower with 0.48s, missing the dead-

line for the shortest distance (0.45s) and therefore

failing to resolve the conﬂict in the third and clos-

est run at 15m.

5 DISCUSSION

The purpose of this research was to reach a resolution

in the ﬁrst place in the event of a conﬂict. The results

in section 4.3 show that in either case, whether the

96%

92%

12%

0% 20% 40% 60% 80% 100%

Competing

Accommodating

Avoiding

Decidability

78%

67%

100%

0% 20% 40% 60% 80% 100%

Competing

Accommodating

Avoiding

Correctness

0.48s

0.1s

0.08s

0 0.2 0.4 0.6 0.8 1

Competing

Accommodating

Avoiding

Time

Deadline 15m

0.45s

Deadline 25m

0.75s

Figure 10: Deadline met before traversing whole tree.

algorithm managed to fully traverse the decision tree

or not, it managed to resolve the conﬂict in 97% of all

the scenarios generated.

Leaving 3% of scenarios in which the scenarios

were near identical and there were basically no differ-

ences between the two lanes; at this point, choosing

one or the other randomly would not make a differ-

ence, given that the analyses conducted were unable

to distinguish between them.

5.1 Comparing Different Techniques

The Competing technique provides the best result of

all techniques, reaching a resolution in most cases

while excelling in situations where the algorithm

couldn’t traverse the entire decision tree. Having a

decidability of 96% and an accuracy of greater than

75%, covers situations where there was insufﬁcient

time to traverse the entire tree. The main drawback

is that it is the slowest, as it takes longer to check all

the leaves of the decision tree to determine which lane

could have earned the maximum potential score.

The Avoiding technique provided the most accu-

rate result, with an accuracy of 100% in both cases,

As if the decision tree traversal had not been termi-

nated, the decision tree scores would not have been

conﬂicting in the ﬁrst place, with one of the lanes hav-

ing a higher priority than the other. Furthermore, It’s

also the fastest among them all. The main downside is

that, only 11% of scenarios are decidable on average.

The Accommodating technique had the highest

decidability in both cases, at 93%, ensuring a reso-

lution if the others failed. However, it had the low-

est correctness, with an average of 66%; this percent-

age would increase as the self-driving car encountered

more roads and scenarios, as it is the only technique

that relies on the history log. Therefore, the more data

that is available, the more precise the results will be.

In conclusion, the three techniques excel in dis-

tinct aspects. Therefore, integrating the results of all

Conﬂicting Moral Codes for Self-Driving Cars: The Single Car Case

255

three techniques to enable the self-driving car to

make a decision would yield the best possible results.

6 CONCLUSIONS & FUTURE

WORK

The aim of the paper was to resolve conﬂicts that

could arise in a self-driving car in the event of un-

avoidable accidents. A conﬂict resolution technique

was implemented using the Thomas-Kilmann Con-

ﬂict Model, along with the Verifying Nodes technique

to resolve the conﬂicts that may arise. Both tech-

niques were applied over a decision tree that was con-

structed based on the MIT Moral Machine Experi-

ment results (Awad and Dsouza, 2018). We managed

to create a system where we tested 80 different con-

ﬂicting scenarios achieving a decidability of 97% in

the scenarios, and an average accuracy of 80%. Over-

all, we managed to resolve 97% of the conﬂicting sce-

narios generated, in an average of just 0.2 seconds,

leaving out only 3% where the scenarios were fairly

identical that randomly choosing one over the other

would have made no difference.

Future research can go in several directions. First,

the presented testing scenarios can be enriched by

gathered more data about the environment using im-

age processing techniques. The behaviour of the pro-

posed conﬂict-resolution model can then be further

veriﬁed. It is also worth investigating the combina-

tion of the three different resolution techniques which

could potentially result in faster and even more ac-

curate results. Extending the work presented in this

paper beyond the single car case to resolve conﬂicts

among a swarm of cars is also a natural next step.

REFERENCES

Anbar, D. (1983). On estimating the difference between two

probabilities, with special reference to clinical trials.

Biometrics, pages 257–262.

Awad, E. and Dsouza (2018). The moral machine experi-

ment. Nature, 563(7729):59–64.

Barabas, I., Todorut¸, A., Cordos¸, N., and Molea, A. (2017).

Current challenges in autonomous driving. In IOP

conference series: materials science and engineering,

volume 252, page 012096. IOP Publishing.

Bazan, J. G., Bazan-Socha, S., Buregwa-Czuma, S., Dydo,

L., Rzasa, W., and Skowron, A. (2016). A classiﬁer

based on a decision tree with verifying cuts. Funda-

menta Informaticae, 143(1-2):1–18.

Borenstein, J., Herkert, J., and Miller, K. (2017). Self-

driving cars: Ethical responsibilities of design en-

gineers. IEEE Technology and Society Magazine,

36(2):67–75.

Cowger Jr, A. R. (2018). Liability considerations when au-

tonomous vehicles choose the accident victim. J. High

Tech. L., 19:1.

Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., and

Koltun, V. (2017). Carla: An open urban driving sim-

ulator. In Conference on robot learning, pages 1–16.

PMLR.

Fahy, A. S., Mueller, C., and Fecteau, A. (2021). Con-

ﬂict resolution and negotiation in pediatric surgery.

In Seminars in Pediatric Surgery, volume 30, page

151100. Elsevier.

Huber, P. J. (1965). A robust version of the probability ra-

tio test. The Annals of Mathematical Statistics, pages

1753–1758.

Junietz, P., Bonakdar, F., Klamann, B., and Winner, H.

(2018). Criticality metric for the safety validation of

automated driving using model predictive trajectory

optimization. In 2018 21st International Conference

on Intelligent Transportation Systems (ITSC), pages

60–65. IEEE.

Karnouskos, S. (2018). Self-driving car acceptance and

the role of ethics. IEEE Transactions on Engineering

Management, 67(2):252–265.

Lin, P. (2016). Why ethics matters for autonomous cars. In

Autonomous driving, pages 69–85. Springer, Berlin,

Heidelberg.

Safavian, S. R. and Landgrebe, D. (1991). A survey of de-

cision tree classiﬁer methodology. IEEE transactions

on systems, man, and cybernetics, 21(3):660–674.

Schaubhut, N. A. (2007). Thomas-kilmann conﬂict mode

instrument. CPP Research Department.

Tamke, A., Dang, T., and Breuel, G. (2011). A ﬂexible

method for criticality assessment in driver assistance

systems. In 2011 IEEE Intelligent Vehicles Sympo-

sium (IV), pages 697–702. IEEE.

Wagner, S., Groh, K., Kuhbeck, T., Dorfel, M., and Knoll,

A. (2018). Using time-to-react based on naturalistic

trafﬁc object behavior for scenario-based risk assess-

ment of automated driving. In 2018 IEEE Intelligent

Vehicles Symposium (IV), pages 1521–1528. IEEE.

Williams, G. O. (1996). The use of d’as a decidability index.

In 1996 30th Annual International Carnahan Confer-

ence on Security Technology, pages 65–71. IEEE.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

256