Reinforcement learning and multi-agent systems-driven optimization of emotional perception and collaborative response mechanisms in intelligent customer service systems

Intelligent customer service platforms depend on effective emotion recognition and adaptive response generation to enhance user experience and service efficiency. Conventional dialogue systems are generally built on rule-based strategies or single-agent reinforcement learning techniques. Such approaches often encounter difficulties in managing extended conversations, adapting to fluctuating emotional states, and coordinating decisions across multiple interacting agents. To overcome these limitations, this study introduces an emotion-aware Multi-Agent System (MAS) designed to generate appropriate responses while maintaining stability and operational effectiveness. The proposed Multi-Agent Soft Actor-Critic with Grizzly Bear Fat Increase Optimization (MASAC-GBFIO) combines multi-agent reinforcement learning and bio-inspired optimization to strengthen learning capability and cooperative agent behavior. Experiments were conducted using the Emotion-Aware Multi-Agent Dialogue Dataset, which contains 6,500 records representing customer interactions from various service domains. Data preparation involved handling missing values, normalizing attributes, and tokenizing textual content to improve data reliability. Sentence-BERT (SBERT) embeddings were employed to capture semantic meaning, contextual relationships, and emotional characteristics within dialogues. The MASAC-GBFIO structure incorporates dedicated agents responsible for emotion analysis, policy formulation, knowledge retrieval, and escalation handling. Multi-Agent Soft Actor-Critic supports coordinated learning among agents, whereas Grizzly Bear Fat Increase Optimization refines hyperparameters and reward scaling to improve adaptation. Implemented in Python, the proposed system outperformed ESDP and GPT-2, achieving a response time of 1.96 s, a BLEU-4 score of 42.45, and a ROUGE-L score of 64.59. Results demonstrate the potential of integrating emotion-aware multi-agent reinforcement learning with bio-inspired optimization for responsive and context-sensitive customer service dialogue management.