Imitation Learning Based On Generative Adversarial Network

Machine learning uses data or past experience to improve the performance of computer programs.It is considered as an important way to achieve artificial intelligence,and has received a lot of attention and extensive applications in the fields of computer vision,data mining,and natural language processing.With data sampled from expert guidance,imitating learning can get a direct reference to each step of decision and learn the end-to-end strategy model.Thus,imitation learning has received extensive attention from researchers in recent years.The generative adversarial networks(GAN)has become a hot research direction in the field of artificial intelligence.We will use the generative adversarial networks to realize the imitation learning,avoid the acquisition of reward function and realize the end-to-end learning from expert trajectory to strategy.We mainly analyze the implementation ideas of traditional imitation learning algorithms,exploit the metrics that can uniquely determine the strategy,and give the feasibility of implementing imitation learning through the metrics.We exploit the consistency of the generated model objective function and the simulated learning objective function in the generative adversarial networks,realize the optimization process of the imitation learning algorithm by the combination of the adversarial network and the generative network.Thereby imitation learning is linked with the generative adversarial networks.The traditional generation of GAN has certain defects.We analyze the characteristics and reasons of defects,then uses the WGAN to realize imitation learning.By determining the new objective function,the imitation learning algorithm can avoid the defects caused by traditional GAN.We discusses the update process of the strategy model in the process of training network,adopting the algorithm that can update the strategy model.We uses OpenAI's simulation physics engine Mujoco as a simulation platform.By performing simulation learning algorithms on multiple tasks provided in Mujoco,this paper verifies the feasibility of using the GAN idea to realize imitation learning and realize the end-to-end learning from expert trajectory to strategy.