An approximate process for designing ethical environments with multi-agent reinforcement learning

Arnau Mayoral Macau, Manel Rodriguez-Soto, Maite López-Sánchez, Juan Antonio Rodriguez Aguilar, Enrico Marchesini, Alessandro Farinelli

Keywords: Value alignment, Multi-agent reinforcement learning, Deep Reinforcement Learning, Multi-objective reinforcement learning, Ethics

TL;DR: In this work, we tackle the value alignment problem by presenting a novel Deep Multi-Agent Reinforcement Learning algorithm called AMAEEP, for designing environments where agents are incentivised to learn to behave ethically.

Abstract: This paper presents an algorithm for designing an environment where multiple autonomous agents learn to behave aligned with a moral value while pursuing their individual objectives. Based on the Multi-Objective Reinforcement Learning and Deep Reinforcement Learning literature, our algorithm represents an extension of the Multi-Agent Ethical Embedding Process (MAEEP), a theoretically grounded method that can be just applied to small problems. We call our method Approximate Multi-Agent Ethical Embedding Process (AMAEEP) and empirically evaluate it in an ethical extension of the gathering game that considers the value of beneficence. Although this environment is much larger than the one used to illustrate the application of the original MAEEP, our method succeeds in dealing with the complexity increase.

Type Of Paper: Full paper (max page 8)

Anonymous Submission: Anonymized submission.

Submission Number: 24