Automated Ethical Design of Multi-Agent Reinforcement Learning Environments

Abstract

This paper introduces the Approximate Multi-Agent Ethical Embedding Process, an algorithm to ethically design reinforcement learning environments where agents learn behaviours aligned with a moral value, while pursuing their own goals. Building on Multi-Objective and Deep Reinforcement Learning, it extends a previously theory-driven method limited to small-scale problems. The new approach is tested in a scaled-up, ethically augmented version of the gathering game, demonstrating its effectiveness in managing increased complexity.