Research on Multi-NPC Marine Game AI System based on Q-learning Algorithm

As a typical representative of human-machine intelligence, games have always pursued the fidelity of NPCs to humans. In the traditional game design process, the behavior of NPCs is usually solidified, and a fixed behavior tree is designed for each NPC so that NPCs can choose appropriate behaviors according to the actual game state during the game progressing process. However, the behavior tree of this fixed coding mode is rigid in behavior and needs to design a complex and huge behavior tree to describe the complex NPC logic. As a dynamically upgradeable behavior tree, Q-learning has strong online adaptability. The artificial intelligence AI behavior tree is designed through the Q-learning algorithm, which can evaluate the environment and action selection and reversely affect the action plan to make human-like behavior feedback. This paper optimizes the Q-learning algorithm and proposes an optimized NPC behavior tree starting from the idea of simulated annealing.