Abstract: The traditional parameter optimization algorithm cannot meet the requirements of the application scenarios with high real-time, large amount of calculation. Combining with the current mainstream machine learning methods, an improved parameter optimization method based on back-propagation BP neural network is proposed. Using the data of the local search algorithm to train the network and predict the parameters, the proposed methods can obtain better real-time performance and lower computational complexity for the traditional search algorithm is replaced. The proposed method is compared with random forest and XGBoost methods. The simulation results show that the order of magnitude of the mean square error of each parameter predicted by BP neural network is 10^-6 or less. The average value of the ratio between the key generation rate calculated by the predicted parameter and the optimal key generation rate is 0.9988. And BP neural network in this application has better prediction performance than random forest and XGBoost.