Design And Implementation Of Non-player Characters AI System Based On Reinforcement Learning

The field of Artificial Intelligence(AI)is developing rapidly.The reinforcement learning algorithm has been well applied in many fields.With the perceived ability of deep learning method,deep reinforcement learning has become a way to solve many real-world problems.In recent years,game industry is gradually maturing in China.Consumers have higher requirements for the quality of games.All this requires game developers to spend more time on game development and the gameplay improvement to meet the needs of consumers.The Non-player characters(NPC)is an important part of most types of games.The performance of NPC AI can directly affect the quality of the game.Developers need a new NPC AI design technique with low complexity and good extensibility.Therefore,this thesis provides developers with a new approach to developing the NPC AI.This thesis designs an NPC AI development system based on proximal policy optimization(PPO)algorithm.According to the characteristics of the NPC in games,the system trains the intelligent agent by PPO as well as improves the effect of the algorithm by adding an additional policy network and using the experience replay method.The framework of the system is based on C/S Architecture.This achieves the information interaction between the system and the game environment so as to complete the model training.The system is using a hierarchical architecture for reducing the dependency.The core policy layer of the hierarchy is mainly implemented by PPO,and the easy-to-use API increases system flexibility.Experimental test proves that the proposed NPC AI system has a good performance in character intelligence.Comparing with the mainstream methods of the industry,the system has a significant improvement in difficulty and extensibility.The system can be applied to various types of video games during the development to help the developers in saving costs.