Emotional Preferences Regulate Goal Priorities in Reinforcement Learning Agents

A fundamental challenge in designing intelligent agents is determining how they prioritize competing objectives, especially when external environments and internal states are constantly changing. This research explores whether emotional preferences, autonomously generated by higher-level goals, can dynamically regulate these lower-level objective priorities, rather than relying on pre-specified settings. Inspired by the goal-directed theory of emotion, the paper introduces a computational framework built upon multi-objective reinforcement learning. This framework consists of an inner controller that provides a range of preference-conditioned behaviors and an outer preference generator. The outer generator learns to map current states to objective preferences through reinforcement learning, driven by a high-level goal. This approach operationalizes emotional preference as a state-dependent regulation of goal priorities that emerges through optimization. Experiments in multi-objective exploration environments show that the learned preference function successfully exhibits contextual priority switching, graded trade-offs, and temporal persistence. This dynamic regulation significantly outperforms strategies based on fixed or handcrafted preferences.