Application and design of a decision-making model in ethical dilemma for self-driving cars

Abstract

Artificial intelligence (AI) has promoted application and development of self-driving cars. However, when self-driving cars encounter ethical dilemma, it is still hard to make a satisficing and clear decision-making by these present moral rules and mechanisms, which makes people distrust in self-driving cars in real life. It is necessary to design a computational and multi-factor decision-making model for self-driving cars. ACWADOE (WADOE Based on Attribute Correlation) is proposed to achieve probabilities of going straight and swerving in ethical dilemmas from more influencing factors to make satisficing and clear decision-making as far as possible. In order to construct ACWADOE model, the prior probability between influencing factor and decision-making is calculated by survey data in moral machine, which can express human preferences and tendencies, align with the requirements of the majority. Then 116 dilemmas are designed and chosen to solve correlation coefficient between influencing factors. Moreover, 84 comparative dilemmas are designed to achieve information gain ratio between influencing factors, then the weight of each factor in decision-making can be calculated by constructing pairwise comparison matrix. Lastly, 40 dilemmas are used to test and verify NB (Naive Bayes), ADOE (Averaged One-Dependence Estimators), WADOE (Weighted ADOE) and ACWADOE respectively. The test results show that ACWADOE is more suitable with human requirements than other models, its accuracy is 92.5%. Furthermore, ACWADOE not only provides a computational decision-making model in ethical dilemma for self-driving cars, but also provides a few references for other AI systems to solve ethical dilemma, which is conducive to make satisficing and clear decisions.