Implementação e análise de perfis éticos em veículos autônomos

Abstract

The advancement of artificial intelligence has introduced autonomous vehicles into dynamic urban scenarios, raising crucial questions about decision-making in high-risk situations. The mere application of strict traffic regulations is insufficient to handle complex moral dilemmas, such as imminent accidents involving pedestrians or multi-vehicle collisions, demanding systems capable of weighing rules and consequences. In this context, this work aims to develop a simulation environment to implement, analyze, and compare different ethical profiles in autonomous vehicles using the BDI architecture and Reinforcement Learning. The methodology consisted of creating a two-dimensional simulator in Python, using the Pygame library to reproduce scenarios of signalized intersections and highways. Vehicles and pedestrians were modeled as autonomous agents based on the BDI (Beliefs, Desires, and Intentions) architecture through the MASPY framework. Five moral behavior profiles were established: Deontological, Legalistic, Utilitarian, Egoistic, and an RL Learning Agent. While the purely symbolic agents acted based on predefined logical utility rules, the learning agent employed Reinforcement Learning techniques (Q-Learning) to adapt its logical decisions based on punishments and rewards linked to the preservation of life, compliance with laws, and traffic flow efficiency. The system ensured the automatic extraction of critical metrics in JSON format, including pedestrian collision rates, vehicle collisions, traffic infractions, and average waiting times. The simulations demonstrated the technical feasibility of incorporating moral pluralism into the deliberation of multi-agent systems, highlighting the operational and ethical impacts that each profile produces on road safety. It is concluded that the union of symbolic and empirical methods contributes significantly to the development of more flexible and transparent computational agents aligned with human protection in autonomous mobility.