Abstract

Game agents in open environments often struggle with traditional exploration due to sparse environmental rewards. This study adopts deep reinforcement learning to enhance agent decision-making in reward-deficient electronic game settings. We developed a human-machine collaboration model that utilizes natural language instructions to guide the reinforcement learning process through reward construction. To address the sparse feedback problem, Hindsight Experience Replay (HER) was integrated into the architecture. Experimental results show that the natural language reward model achieved a 92% prediction accuracy and a game score of 9.8. Following HER optimization, target instruction accuracy reached 97.8% with a final score of 9.9. These findings demonstrate that combining linguistic guidance with experience replay significantly improves application performance in coefficient reward environments.