Ethical Decision-making for the Inside of Autonomous Buses Moral Dilemmas

Abstract

The emergence of moral dilemmas inside autonomous buses can lead to unpredictable consequences for maneuver and amplify potential harms to passengers onboard. However, most existing approaches solely focus on ensuring ethical decision-making in scenarios outside of vehicles. To ensure ethical decision-making for autonomous buses when moral dilemmas occur inside, there are many urgent challenges that need to be addressed. First, the noncommensurability of ethical values presents difficulties in designing quantifiable environments and decision-making models. Moreover, ethical dilemmas involve multiple conflicting objectives, often necessitating the consideration of multiple moral theories to comprehensively evaluate different perspectives. Additionally, accurately representing these dilemmas and identifying optimal solutions that address conflicting objectives poses further challenges. This article proposes a general ethical decision-making system to handle ethical dilemmas inside autonomous buses. The system’s design adheres to multiple ethical principles, and it comprises two stages: 1) develop a generative adversarial network (GAN) based human-value-aligned data collection scheme to gather representative moral values and generate comprehensive moral scenarios, which address the incommensurable ethical metrics issue; and 2) propose an ethical compliant multiobjective thresholded lexicographic Deep Q-learning method to ensure optimal policies that satisfy multiple ethical objectives. In a case study of autonomous bus route planning, our system outperforms benchmarks in showing a greater number and more evenly distributed policies in the Pareto Front and a 27% increase in coverage rate. Extensive experiments against nonethical systems show superior outcomes in convergence, average reward and cost. Finally, user studies demonstrate the system’s accuracy and usability.