Ethical Decision-Making in Autonomous Multi-Agent Systems for Disaster Relief : An Ethical AI Approach to Autonomous Disaster Response

Sammanfattning

In times of disaster, rapid decision making and efficient coordination can mean the difference between life and death. Multi-Agent Systems (MAS), powered by Artificial Intelligence, are increasingly being explored for their ability to handle the complex, dynamic nature of disaster relief-deploying autonomous agents to search for victims, deliver supplies, or assess damage. However, as these systems become more autonomous, they raise critical ethical questions: Are the decisions fair? Are they transparent? Can we trust them? This thesis addresses these concerns by developing a Multi-Agent Reinforcement Learning (MARL) framework that doesn’t just focus on performance, but puts ethics—specifically fairness, transparency, and reliability—at its core. Drawing inspiration from the European Union Guidelines for Trustworthy AI, the framework is designed to ensure that autonomous agents operate not only efficiently, but also in a way that aligns with humanitarian principles. To do this, we create a simulated disaster environment where agents are trained using a reward function that incorporates three ethical metrics - Fairness, Transparency and Reliability. Fairness is measured by how equitable resources are distributed, considering victim severity, urgency, and even distance. Transparency reflects how confident the agents’ decisions are, while reliability ensures that similar situations lead to consistent actions. The study builds a structured ethical dataset and an evaluation toolkit to assess how well agents adhere to these principles in practice. By comparing our ethically guided model to traditional performance-driven approaches, we show that it is possible and necessary to design AI systems that act with both intelligence and integrity. This research contributes a step forward in building AI that we can trust, especially when lives are at stake.