Human-in-the-loop Reinforcement Learning for Emotion Recognition

Facial emotion recognition (FER) analyses a person's facial expressions in images to understand the person's emotions. Achieving high accuracy in real-world FER systems requires extensive data, which poses challenges in cost, practicability, and privacy, especially in ensuring fairness and avoiding discrimination related to skin color or health conditions. This paper introduces a novel approach called two-state Q-learning with human feedback (TS-QL-HF) to enhance the accuracy of a FER system by integrating human feedback collected during the evaluation process as a refined reward function into two-state Q-learning (TS-QL) and double Q-learning (DQL) to operate in deterministic environments. This approach does not need an extensive dataset to include minority populations. Simulation results demonstrate that TS-QL-HF provides the highest accuracy. However, the tuned TS-QL, without human feedback, shows a higher efficiency despite a slightly lower accuracy due to the inaccurate reward function.