Extraction and Discussion of Influencing Factors in Self-driving Ethical Decision

doi:10.21203/rs.3.rs-1960111/v1

Download PDF

Article

Extraction and Discussion of Influencing Factors in Self-driving Ethical Decision

https://doi.org/10.21203/rs.3.rs-1960111/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

At present, many scholars found many influencing factors in self-driving ethical decision by the way of questionnaire and theoretical researches. However, the important influencing factors can’t still be extracted and determined in self-driving ethical decision, which is not conducive to construct decision model and framework in moral dilemmas and causes information overlap and interference by multi-variable and multi-collinearity. Therefore, it is necessary to extract a few important factors from these influencing factors in self-driving ethical decision. 116 normal dilemmas and 10 comparative dilemmas, in which there are only two influencing factors to be compared, are designed to be surveyed in this paper. According to the probability of the straight choosing as protector, the comparative result and location in decision tree model, the importance is determined in self-driving ethical decision, then a few important influencing factors are extracted and determined in self-driving ethical decision. Eventually, these factors of species, harm, traffic laws, number, age and self-interests are regard as the important influencing factors in self-driving ethical decision, which provide a few theoretical guidance to construct and design model and framework in self-driving ethical decision for car manufactures. Furthermore, they provide a few references to formulate traffic laws and regulations related to self-driving technology for traffic management authentication.

Self-driving technology

Ethical decision

Influencing factor

Factor extraction

Moral dilemma

With development of artificial intelligence, self-driving technology has made great progress and a few car manufactures have produced self-driving car with L2 or L3 level¹. While self-driving cars, like trams, suffer from a kind of trolley dilemma, which refers to the fact that there are two or more options during decision process in a specific scenario, and these options are related to each other and there are also a few contradictions between them. it must be hard for self-driving car to choose a reasonable option². As an independent decision-maker in the future, self-driving cars should make decision in moral dilemmas. However, human being is hard to make fair and reasonable decision when faces moral dilemma, not to mention that self-driving cars can make reasonable and right decision in moral dilemmas³. Therefore, it is necessary to strengthen research on self-driving decision in moral dilemmas. In China, Commission for Deepening all-round Reform of the CPC Central Committee deliberated and approved some guiding opinions on strengthening ethical governance in science and technology, which indicates that Chinese government has made a top-level design for strengthening ethical governance in science and technology⁴. Only by strengthening ethical governance of science and technology can we shape cultural concept and guarantee mechanism of science and technology to be good. Because science and technology is a “double-edged sword”, it may let science and technology go along right direction and promote human economic and social development, benefit mankind, and avoid negative consequences of science and technology development on human survival and development. In order to study self-driving ethical decision very well, it is necessary to understand how human beings make decision in moral dilemmas in advance⁵. Only in the way can we construct a self-driving ethical decision model, which is suitable for human morality to improve human acceptability and trust in self-driving car. Therefore, many research institutions and scholars have studied influencing factors of decision-making in moral dilemmas by questionnaire and experiments. For example, Moral Machine, an online experiment platform launched by Massachusetts Institute of Technology in 2016, aims to explore how human beings make decisions when face moral dilemmas. How do human beings assess moral preferences by understanding what human beings want self-driving car to do in unavoidable accident to solve moral dilemmas. The survey results, published in Journal of Nature in October, 2018, indicated three strong preferences: to protect human, to protect more lives and to protect young⁶. Anja K. Faulhaber et al. explored impact of quantity, age, disability and context on human decision in moral dilemmas by virtual driving platform. The experiment results indicate that human are more likely to protect more and younger. The context has little impact on decision and people with disability is given greater priority under the same condition. When the number of pedestrians is more 4–5, the user or owner in self-driving car is willing to sacrifice themselves to protect more pedestrians on the road⁷. Bonnefon et al. found by experiments that users or owners in self-driving cars are more willing to protect the passengers in self-driving car than pedestrians on the road based on deontology⁸. James et al⁹ found that human choice was different when respondents knew and didn’t know injury consequence in an accident. Under the condition of not knowing injury consequence and there was only a passenger in self-driving car, 92.4% respondents chose to collide with the static obstacle in the front. When there were two passengers, only 86.6% respondents chose to collide with the static obstacle. It indicates that the number has an important influence on self-driving ethical decision. When respondent knew that the passengers in self-driving car were seriously injured in an accident, 97.9% respondents chose to collide with pedestrians on the road, which indicates that harm and self-interest have greater impact on self-driving ethical decision than number in a serious accident. Bergman et al¹⁰ also demonstrated choices in moral dilemmas by test. The test results show that age has more impacts on public choice. 90% of testers choose to save child and 72% choose to save adult between adult and the elderly. More testers will choose to save the younger. Darius Aurel Frank et al¹¹ investigated that choice is various from different perspective. For example, more respondents, as passenger, are likely to choose pedestrians to be sacrificed than those as pedestrian or bystander. In addition, when respondents’ choice time and way are different, their results are also different. For example, when respondents were asked to make choice in an intuitive way and within 5 seconds, more respondents would choose to protect pedestrians on the road. If respondents make choice in a rational way and for a long time, less respondents would choose to protect passengers in self-driving car. Celso et al¹² found that more respondents chose utilitarian principle to protect 5 pedestrians on the road than 1 passenger in self-driving car in tunnel dilemma. However, most respondents are still unwilling to buy self-driving cars with utilitarian in reality. the theoretical connotation of human decision should be combined with practical guiding principles in designing intelligent machine as far as possible to solve moral dilemma. Furthermore, regional differences in culture, customs, concepts and laws can also impact on self-driving decision in moral dilemmas¹³. Therefore, there are too many influencing factors in self-driving ethical decision. But all influencing factors are considered during decision process, which results in information overlap and interference caused by multi-variable and multi-collinearity¹⁴. the importance of each influencing factor should be determined in this paper by analyzing survey data in normal and comparative moral dilemmas and location in decision tree model, then a few major influencing factors would be extracted and determined, which provide theoretical guidance to establish self-driving ethical decision model and algorithm in the future.

In order to extract a few major influencing factors in self-driving ethical decision, 116 normal moral dilemmas are designed and surveyed by online and offline questionnaires. 82 moral dilemmas are surveyed by online way, which requires each respondent randomly choose 15 dilemmas from 82 to be surveyed. A total of 844 valid questionnaires are received and 12660 data items are collected. All respondents are adults over 18 years old, 321 are female. Other 34 dilemmas are surveyed by on-site questionnaire. 162 respondents who are students and teachers in an university, fill in questionnaires on the spot, 5508 data items are collected. Then these survey data is analyzed and counted from gender, number, age, social status, traffic laws, harm, and species. .

In terms of gender, if self-driving car chooses to go straight and the injured are female, 2843 copies choose the straight as protector and 1656 as sacrificer in Table 1. Protection rate is 63.19%. However, if the injured are male in the straight, 1154 copies choose the straight as protector and 1377 as sacrificer. Protection rate is 45.39%. the difference between both protection rates is 17.8%.

For the influencing factor of number, when self-driving car chooses to go straight and injuries more road users in an accidence, 2763 copies regard the straight as protector and 1739 as sacrificer. Protection rate is 61.37%. On the contrary, the injured road users are less in the straight, 1647 copies choose the straight as protector and 2225 as sacrificer. Protection rate is 42.54%. the difference between both is 18.83%, as shown in Table 2.

For age, when pedestrian in the straight is younger than that in the turning, 1716 copies choose the straight as protector and 1147 as sacrificer. Protection rate is 59.94% for the young. However, pedestrian in the straight are older, 1453 copies choose the straight as protector and 2065 as sacrificer. Protection rate is 41.3% for the elder, the difference between both protection rate is 18.64% in Table 3.

About social status, when people in the straight have higher social status, 2288 copies choose the straight as protector and 1587 as sacrificer. Protection rate is 59.04%. Otherwise, when people with lower social status are in the straight, 1614 copies choose the straight as protector and 2267 as sacrificer. Protection rate is 41.59%. The difference between both protection rates is 17.45%, as shown in Table 4.

Considering traffic laws’ influence on decision, when pedestrians in the straight obey traffic laws, 4353 copies choose the straight as protector and 2485 as sacrificer. Protection rate is 63.66%. On the contrary, when pedestrians in the straight violate traffic laws, 807 copies regard the straight as protector and 1862 as sacrificer. Protection rate is 33.42%, The difference between both protection rates is 30.24% in Table 5.

In terms of harm, human is concerned about whether decision behavior will cause harm to pedestrians or passengers. Harm has an important role on self-driving ethical decision. When self-driving car choose to go straight, there are more harms to road users in the straight. 1613 copies choose the straight as protector and 613 as sacrificer. Protection rate is 72.46%. On the contrary, there are less harms in the straight than in the turning. 159 copies choose the straight as protector and 165 as sacrificer. Protection rate is 49.07%. the difference between both protection rates is 23.39% as shown in Table 6.

When road user in the straight is mankind and that in the turning is animal or inanimate, 1084 copies choose the straight as protector and only 296 as sacrificer. Protection rate is 78.55%. Otherwise, road user in the turning are mankind and that in the straight is animal or inanimate, 169 copies choose the straight as protector and 227 as sacrificer. Protection rate is 42.68%, the difference between both is 35.87%. The statistical result is in Table 7.

In a result, the more the difference between both protection rates for each influencing factor, the greater the influence on self-driving ethical decision. The differences between both protection rates for the 7 influencing factors are shown in Table 8.

As it can be seen from the above table, species has the greatest impact on self-driving decision in moral dilemmas, social status and gender have less impact among 7 influencing factors.

Considering these influencing factors including species, number, age, traffic laws, social status, gender and harm, 10 comparative dilemmas are designed to compare influence on ethical decision among these influencing factors. A few college students and teachers are selected as respondents by online survey. A total of 616 valid questionnaires are received, more 6000 data items are collected.

The first moral dilemma is used to compare influence between number and age in Fig. 1. A child is crossing in the straight and two adults are walking in the turning. If self-driving car chooses to go straight, the child is injured. Otherwise, self-driving car chooses to swerve, two adults are injured. According to survey result, 92.05% of respondents choose to go straight and only 7.95% choose to swerve, which indicates that number has greater influence on self-driving ethical decision than age.

For comparative dilemma between age and traffic laws, just as shown in Fig. 2. a little girl is crossing the road without obeying traffic laws in the straight and an old lady following traffic laws is crossing in the turning. Any action self-driving car takes will result in death of the little girl or the old lady, how to choose for human? 86.36% of respondents choose to go straight and 13.64% choose to turn in moral dilemma, which indicates that traffic laws has greater influence than age.

Which factor has a greater impact between age and harm? A comparative dilemma between age and harm is designed in Fig. 3. A man is in a self-driving car. When self-driving car goes straight, it collides with a static barrier in the front, which cause slight harm to the passenger in self-driving car. Otherwise, the self-driving car choose to swerve, it will lead to death of the elderly crossing the road. According to human survey, 88.64% of respondents choose to go straight to collide with the barrier and only 11.36% choose to swerve, which indicates that harm has greater impact than age.

For species and harm comparison just as shown in Fig. 4. there is a static barrier in the straight and a cat crossing the road in the turning. When a self-driving car with an old lady chooses to go straight to collide with the barrier, the old lady will be slightly injured. If the self-driving car swerves, the cat will be killed. How about human choice? The survey results showed that 36.36% of respondents chose to go straight and 63.64% chose to swerve, which demonstrates that species has greater influences than harm in self-driving ethical decision.

For species and traffic laws, number comparative dilemma just as shown in Fig. 5. An adult in the straight is crossing the road at a red light and two dogs are following traffic laws to cross the road in the turning. If the self-driving car chooses to go straight, the adult will be killed. Otherwise, swerving will kill two dogs. The survey results shows that 34.09% of respondents choose to go straight and 65.91% choose to turn to kill two dogs, which indicates that species has more influence than traffic laws, number.

Which makes greater influence on ethical decision between harm and traffic laws? There is a moral dilemma in Fig. 6. there is an old passenger in a self-driving car. When the self-driving car goes straight, it hits a barrier and the passenger will be slightly injured. If the self-driving car swerves, an old pedestrian crossing at a red light will be killed. The survey results indicates that 55.68% of respondents choose to go straight and 44.32% choose to swerve, which demonstrates that harm has greater influence than traffic laws.

In terms of harm and number comparison just as shown in Fig. 7. when a self-driving car goes straight to collide with a static barrier in the front, three passengers in the car will be slightly injured. Otherwise, swerving will kill a pedestrian on the road. According to human survey, 75% of respondents choose to go straight and 25% choose to swerve, which indicates that harm has greater influence than number.

For traffic laws and number comparison, two pedestrians are crossing in the straight, who run a red light in the dilemma. At the same time, a pedestrian is crossing at a green light in Fig. 8. A self-driving car goes straight to kill two pedestrians in the straight. Otherwise, it will kill one pedestrian in the turning. How to choose for human? The survey results show that 80.64% respondents choose to swerve and 19.36% choose to go straight. It demonstrates that number has greater influence than traffic laws in a serious accident. However, when pedestrians on both roads are slightly injured, maybe not too many respondents choose to swerve to protect two pedestrians in the straight.

For gender and age in Fig. 9, a little boy in the straight and an old lady in the turning are crossing the intersection. A self-driving car goes straight to kill the little boy, Otherwise, the old lady will die in the turning. According to human survey, 28.39% of respondents choose to go straight and 71.61% choose to swerve to protect the little boy, which indicates age has greater influences than gender.

Which factor is important between age and social status in self-driving ethical decision? A comparative dilemma is designed in Fig. 10. there are a little boy in the straight and a male doctor in the turning. How about human choice? The survey result shows that 66.67% of respondents choose to swerve and 33.33% choose to go straight. It demonstrates that age has more importance than social status.

Based on the above analysis and comparison, it can be seen that species, harm, number, traffic laws and age have greater influence on self-driving decision in moral dilemma than gender, social status et al.

In order to further understand the influence of each factor in self-driving ethical decision, 84 moral dilemmas are selected as training set to construct a decision tree model, which is generated by information gain of each factor. All the factors on common path for any leaf node are important influencing factors in self-driving ethical decision. For information entropy, Shannon introduced entropy into statistical information firstly, then information entropy is also named Shannon entropy¹⁵. Entropy, a common physical concept in physics, is a measure that reflects degree of chaos in a system. The smaller the entropy value is, the more regular the system is and the less information the data carries. On the contrary, the larger the entropy value is, the more chaotic the system is and the more information the data carries^16,17. Supposed X is composed of a few discrete random variables(x₁, x₂,...,x_i), probability distribution of each random variable is p(x = x_i), i = 1,2,...,n, then information entropy of the random variable X is calculated by Formula 1.

H(X) = -\(\sum _{i=1}^{n}{p}_{i}log{p}_{i}\) (1)

conditional entropy refers to uncertain information under a given condition^18,19, which indicates uncertainty of a random variable Y under a known random variable X. It may be shown as H(Y|X), which may be calculated by Formula 2.

H(Y|X) = \(\sum _{i=1}^{n}{p}_{i}\)H(Y|X = x_i) (2)

Information gain refers to the difference between uncertainty of information and uncertainty of information under a condition²⁰, which can be marked as G(D,A). It may be calculated by Formula 3.

G(D,A) = H(D) - H(D|A) (3)

H(D) is information entropy of D, H(D|A) is conditional entropy under known variable A. If information gain corresponding to random variable is larger, the corresponding conditional entropy is smaller, which indicates that the factor has a great influence on self-driving ethical decision. The decision tree model is constructed by ID3.0 and python software, as shown in Fig. 11.

According to the above decision tree model, species is a root node in the decision tree model, so it has the greatest influence in self-driving ethical decision. If species is the same, harm and traffic laws are further considered in decision-making. When harm is serious in the straight, age should be further judged in decision-making. Regardless of age, the party with serious harm is regard as protector. when age has no difference, traffic laws are further considered in decision-making. It indicates that traffic laws have greater influence than age. Furthermore, when species and harm are the same, number is considered in decision-making. In a result, species, harm and traffic laws have more influences than number and age. Moreover, number may be paid more attention to in a serious accident, the more the number, the greater the harm.

From the above analysis and survey, it can be seen that species is a key influencing factor in self-driving ethical decision. No matter any moral dilemma, human beings are given priority to be protected, which conforms to Asimov’s three laws of robot and human morality. For human beings have dual attributes, namely natural attributes and social attributes, and have the capacity of understanding and changing the world. The fact that animals or inanimate cannot means that human beings have higher social value than animals or inanimate and should be protected firstly. Secondly, harm is also a very important influencing factor in self-driving ethical decision. The original goal of developing self-driving technology is to avoid or reduce harm in traffic accidents as far as possible. Although self-driving car cannot fully avoid traffic accidents in reality, it should minimize harm in a accident. Even in order to reduce harm in a accident, self-driving car may be permitted to violate traffic laws. However, traffic laws also play an important role and influence in self-driving ethical decision. Law is a kind of influence of social relation formed between people, which indicates operation of state power and realization of state will, which may be amended, formulated and promulgated by the legislative organ with legislative power to exercise national legislative power in accordance with legal procedures. Only by applying law to the construction and shaping of self-driving ethical decision, can self-driving cars be prevented from replacing legal justice with their own ethical decision and balance their self-decision and legal coercion in ethical decision-making²¹. Lastly, self-driving ethical decision algorithm and mechanism are realized under a legal framework. In the context of current laws and regulations, ethical issues are inevitably be transformed into political pressure and democratic issues in the process of ethical ideas pursuing freedom and independence, which will be reflected in legislation²². In order to realize social justice, law should consciously respond to ethical issues, consider and impact social ethical positions, and obtain its own legitimacy and authority. Just as Hegel said that ‘laws exist for freedom and for themselves, they are rules arising from nature of things’²³. So laws need to be obeyed and enforce in normal circumstances. In view of number, it represents collective scale, safeguards interests of the majority and ensures collective interests. Since collective interests and individual interests are interdependent, only when collective interests are maintained, individual interests can be guaranteed. Therefore, individuals should actively care for and safeguard collective interests. Furthermore, only by putting collective interests above individual interests can it be accepted by human beings. So the factor of number should also be considered in self-driving ethical decision, it is also an important influence factor. Especially in a few serious traffic accidents, self-driving car should choose more road users to be protected. Age is an influencing factor in self-driving ethical decision, which may be against by a few people, especially the elderly, who think it is a discrimination. This behavior is unreasonable and unfair from deontology. However, according to human survey, most respondents usually give priority to protect the younger in an inevitable accident. Especially for children, human beings are willing to protect children, who represent the future and hope. Therefore, age should also be considered in self-driving ethical decision, but it should have less influence than other four factors. In addition, when users or passengers in self-driving car are severely injured and even endanger their lives, they should be give priority to be protected in a accident. Otherwise, self-driving car will become a ghost car, it is hard task to apply self-driving technology in reality.

Through analysis and discussion of survey data and decision tree model in various moral dilemmas, these influencing factors such as species, harm, traffic laws, number and age, as well as self-interests, have important influences in self-driving decision in moral dilemma. Although these influencing factors are extracted from human survey data, there may be some subjectivity and partiality, which cannot meet all people’s morality. But they meet the majority to avoid fatal mistakes in self-driving ethical decision. Furthermore, these influencing factors are extracted and determined in self-driving ethical decision, which provides a theoretical guidance to design and develop decision models and algorithms in self-driving car for car manufacturers. On the other hand, they provides a few references to formulate and amend laws and regulations related to self-driving technology for traffic management authentication.

Data Availability

The data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest

The authors declare no conflicts of interest.

Acknowledgments

This work was supported by the Science and Technology Research Project of Jiangxi Education Department (Grant no. GJJ201911),

1 Zhao Ying. Towards a strong automobile country-Some suggestions on the Mid and Long Term Development Plan of Automobile Industry. Think Tank Theory&Practice, 2(06):63-66(2017).

2 Su Lingyin. The normal dilemma of creating intelligent moral machine and its solving strategy. Theoretical Exploration, 4:30-38(2019).

3 Feng Sumei. Research on self-decision algorithm of AV based on meta-reinforcement learning. Dalian University of Technology. pp:11-17(2020).

4 Commission for Deepening all-round Reform of the CPC Central Committee. some guiding opinions on strengthening ethical governance in science and technology, (2022).

5 Su Lingyin. Construction of ethical framework: The main task in current machine ethics research. Journal of Shanghai Normal University, (01):76-86(2019).

6 Edmond Awad, et al. The Moral Machine experiment. Nature, 563(7729): 59-63(2018).

7 Anja K. Faulhaber et.al. Human decisions in moral dilemmas are largely described by utilitarianism:Virtual car driving study provides guidelines for ADV. Science and Engineering Ethics.(25):399-418(2019).

8 Bonnefon J F, Shariff A, Rahwan I. The social dilemma of autonomous vehicles. Science,352(6293): 1573-1576(2016).

9 James E. Pickering, Mateusz Podsiadly, Keith J.Burnham. A M2D approach for AV normal dilemma:AV collision with a barrier/pedestrian. IFAC Paper online(2019).

10 Bergmann, L. T, Larissa, S, Carmen, M. Autonomous vehicles require socio-political acceptance: An empirical and philosophical perspective on the problem of moral decision making. Frontiers in Behavioral Neuroscience, 12(31): 1-12(2018).

11 Darius-Aurel frank, et al. Human decision-making biases in the moral dilemmas of autonomous vehicles. Scientific Reports, 9:13080-13047(2019).

12 Celso M. de melo, et al. Risk of injury in moral dilemmas with autonomous vehicles. Frontiers in Robotics and AI, 7:572529(2021).

13 Hong Wang, et al. Ethical decision-making platform in autonomous vehicles with Lexicographic Optimization based Model Predictive Controller. IEEE Transactions on Vehicular Technology. 69(8):8164-8175(2020).

14 Wang Xiaoyuan, Yang Xinyue, Wang Fengqun. Research on influencing factors of driving decision based on grey correlation entropy theory. China Safety Science Journal, 17(05):126-134(2007).

15 Wu Yiwen. Research on classification algorithm based on decision tree. Digital Communication World, (12):268-269(2017).

16 Liu Xiaohui. Research on decision tree algorithm based on rough set and ensemble learning. Qingdao University of Science and Technology(2019).

17 Sun Haibo, et al. Unified optimization for multiple active object recognition tasks with attribute decision tree. Journal of Intelligent & Robotic Systems. 103(2): 1- 15(2021).

18 Luigi Vanfretti, V.S. Narasimham Arava. Decision tree-based classification of multiple operating conditions for power system voltage stability assessment. International Journal of Electrical Power and Energy Systems. 123(2020).

19 Asvapoositkul Surat, Preece Robin. Decision tree-based prediction model for small signal stability and generation-rescheduling preventive control[J]. Electric Power Systems Research.196(2021).

20 Bugaje AI, et al. Selecting decision trees for power system security assessment. Energy and AI. 6:21-36(2021).

21 Kumfer W, Burgress R. Investigation into the role of rational ethics in crashes of automated vehicles. Transportation Research Record Journal of the Transportation Research Board, 2489(1):130-136(2015).

22 Brundage, M. Limitations and risks of machine ethics. Journal of Experimental and Theoretical Artificial Intelligence(2014).

23 Kyle Bogosian. Implementation of moral uncertainty in intelligent machines[J]. Minds&Machines, 27:591-60(2017).

Tables 1-8 are available in the Supplementary Files section.

No competing interests reported.

Tables.docx
Tables 1-8

Download PDF

Version 1

posted

You are reading this latest preprint version

Extraction and Discussion of Influencing Factors in Self-driving Ethical Decision

Status:

Version 1

Abstract

Figures

1 Introduction

2 Analysis Of Survey Data In Normal Dilemmas

3 Analysis Of Survey Data In Comparative Dilemmas

4 Analysis By Decision Tree Model

5 Discussion

6 Conclusions

Declarations

References

Tables 1-8

Additional Declarations

Supplementary Files

Status:

Version 1