Intelligent Enhancement Of Non Player Character Based On Deep Reinforce-Ment Learning

Currently,Artificial intelligence is a very hot area of research,the nation has also made AI education as a plan.Perception and decision-makingare very important parts in artificial intelli-gence.Thanks to the rapid development of deep learning,deep reinforcement learning(DRL)has gradually become The key method in this session.DRL uses deep nerual networks to perceive high-dimensional information,extracts key features and take appropriate action based on these input features,then improves the strategy by interacting with the environment.Now,the DRL has become promising in many fields.The field of electronic game has become a benchmark(like atari game series)for testing the DRL algorithms due to its low sample acquisition cost and high speed.The purpose of this thesis is to apply the DRL into the area of game non-player character(NPC)AI design.In this thesis,we make improvement based on deep Q-learning network(DQN).Accordingto the characteristics of NPC's intelligent application,the DQN framework has been improved specifically,and the main contents and contributions are as follows:1.Apply the DRL into the area of game design,especially the NPC AI part.Using the state of the enemys and selves to generate the behavior strategies automatically.On the one hand,by continuous sampling,it can improve the intelligence.On the other hand,automatic gen-eration of behavioral strategies can also greatly improve efficiency.2.Propose a distributed DQN framework.It decouples not only the sampling and learning but also algorithms and agent.Besides multi-process mode accelerates game sampling and learning is of great significance to practical application.Combining priority sampling and imitation learning make it more efficient.3.Propose a hierarchy DQN algorithm(H-ART-DQN).Through training subgoals network to learn a few subgoals and then from the high-level time scale to select subgoals.While speeding up training,different styles of behavior strategies can be generated to cater to game AI design.