Eclectic Ethical Decision Making for Autonomous Vehicles

Ensuring the safety of autonomous vehicles (AVs) in morally ambiguous scenarios is critical for widespread trust and adoption of them. The existing solutions for moral decision-making in AVs are constrained by their methods of handling ethical theories. Specifically, reinforcement learning (RL) approaches generate rewards based on random weights (credence) assigned to different ethical theories, which does not adequately manage uncertainty. Additionally, other approaches like predictive controllers and ethical valence theory rely on preset priorities, which also fail to effectively address the inherent uncertainty in morally ambiguous scenarios. This work proposes a novel approach using RL to managing uncertainty and integrating evidence from multiple ethical theories, aiming to improve the ethical decision-making capabilities of AVs. By combining the simplex algorithm with the Dempster-Shafer (DS) technique, we integrate the theories into our framework to represent and reason with uncertain information. The simplex algorithm optimizes the reward structure in RL, and its results inform the DS technique to generate credence. Our approach outperforms existing models, addressing ethical challenges in AVs effectively.

I. Introduction

Autonomous vehicles (AVs) have shown promise in exceeding human performance in certain areas [1], particularly in reducing traffic congestion and improving safety. Despite these advantages, AVs still face challenges in gaining public trust and widespread adoption [2]. A significant area of ongoing research is decision-making of AVs in ethically ambiguous situations [3]. There is ongoing debate about the most suitable ethical framework for AVs to navigate moral dilemmas [4]. There is a need to explore eclectic strategies for integrating multiple ethical theories within the decision-making process.