Convergence Problem Of Generative Adversarial Network Based On Bhattacharyya Distance

With the improvement of computing power and the development of neural network,more and more scholars devote themselves to the research of generation model.Generative adversarial network,which can be called GAN for short,is a very popular generative model at present.Since the original version of generative adversarial network was proposed in 2014,many researchers have been interested in it.The birth of generative adversarial network is inspired by game theory and adversary thought.The model includes a generator and a discriminator.Through the mutual confrontation,the generator and discriminator are constantly optimized.Finally,make the generator enough to generate fake data.However,there are still many deficiencies in the existing generative countermeasure network model.For example,in the process of training GAN,we need to train the generator and the discriminator.If the discriminator has been trained with high accuracy before the generator can learn enough information of the original probability distribution,it is likely that in the next training,the discriminator will be trained with high accuracy.It is very difficult to get information from the discriminator to optimize itself,which leads to the instability of the training process.In addition,the generative adversarial network is based on neural network,which has the general defect of neural network model,poor interpretability.In addition,many generative countermeasures network models have the phenomenon of pattern collapse,which may repeatedly generate data with little difference.To try to solve these problems,in this paper,a new generation countermeasure network algorithm based on Bhattacharyya distance is proposed by combining theoretical research and algorithm simulation.The existence of the optimal discriminator and its expression form are proved for the designed generation countermeasure network for the arbitrary generator.The global optimal solution of the generator is proved to exist and is just behind the real data It also proves the convergence of the algorithm under ideal conditions and discusses the convergence of the algorithm under non ideal conditions.In addition,we also use pytorch to implement the algorithm and discuss the changes of generator and discriminator in the process of continuous training.The algorithm is also applied to the open source MNIST handwritten digital data set.And discuss the problems in the algorithm and the improved scheme.