A Machine Learning Approach to the Vector Prediction of Moments of Finite Normal Mixtures

The paper compares different architectures of feedforward neural networks designed for vector forecasting of finite normal mixture moments in problems of approximation of plasma density fluctuations. A vector prediction is constructed for mixture expectation, variance, skewness and kurtosis. Its effectiveness is compared to consecutive scalar forecasts for each moment. The best accuracy in terms of RMSE and MAE have been obtained for configurations with one hidden layer of 200 neurons and Adam/Adamax optimizers. In addition, it is demonstrated that neural network effectiveness is increased by using information about other moments without compromising learning speed. This study is focused on the creation of modern computational tools for statistical data processing in important problems of plasma physics.