Aiming at the dynamic scheduling problem of flexible job shop with dynamic insertion of workpieces. an algorithm combined the Large Language Model (LLM) with Deep Reinforcement Learning (DRL) named LLM-DQN was proposed with the objective of minimising the total delay time. By combining the semantic understanding capability of LLM with Deep Q-network (DQN). an integrated framework containing state space optimisation, hybrid action selection and reward function design was constructed. In the state representation. LLM was used to generate weighted feature vectors to highlight key scheduling metrics; in the action selection phase- a hybrid strategy was designed to achieve the dynamic fusion of LLM expert recommendations and DQN strategies; an LLM-driven adaptive reward mechanism was introduced. Simulation experiments on large language models such as DeepSeek and Doubao showed that LLM-DQN outperformed a single scheduling action and other deep reinforcement learning methods in multiple test scenarios.
Wang et al. (Sun,) studied this question.