Title: Navigating the Fog of Love: Shaping Ethical AI Conduct through Affinity-Driven Reinforcement Learning

Abstract: There is growing momentum behind the goal of embedding ethical conduct into artificial intelligence systems. A notable approach in this domain is affinity-based reinforcement learning, a method that employs policy regularization within the objective function to encourage virtuous actions, thereby reducing reliance on the specific design of reward structures. While previous studies have validated this technique in simple scenarios, such as grid worlds and basic toy problems with limited state and action spaces, broader application remains an open challenge. To bridge this gap, we present a two-player multi-agent simulation grounded in the role-playing board game, Fog of Love. Within this framework, two agents strive to balance individual moral goals with the need for relational cooperation. Due to the complexity inherent in this multi-agent dynamic, standard multi-agent deep deterministic policy gradient models have struggled to achieve effective competition or collaboration. Our findings demonstrate that incorporating localized affinities significantly improves agent performance across both competitive and cooperative metrics, as evidenced by higher overall scores. This improvement not only fosters ethical decision-making but also elucidates the agents' underlying purposes, rendering their behavior interpretable at a human level.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC