A meta-learning-based reinforcement learning framework for rapidly adaptive emotion intervention

To address the issue of insufficient personalised adaptability in dynamic emotional interventions, this paper proposes a value-oriented meta-adaptive reinforcement learning framework.By integrating meta-learning and reinforcement learning, the framework constructs a dual-level learning architecture capable of rapidly adapting to individual emotional dynamics with minimal interactions.The model employs a multi-objective reward function to synergistically optimise intervention effectiveness, user engagement, and safety.Experiments on public datasets such as Emotional Support Conversation dataset and Distress Analysis Interview Corpus -Wizard of Oz demonstrate that our approach achieves an emotional state improvement rate of 0.78 and a user satisfaction score of 0.82.These results represent significant improvements over traditional deep reinforcement learning and meta-learning baseline models, providing an effective computational paradigm for addressing adaptability challenges in personalised psychological interventions.