SYNTHESIS OF A STATISTICALLY OPTIMAL DATASET FOR TRAINING A NEURAL NETWORK

To control complex nonlinear systems, intelligent technologies, in particular, based on neural networks, are increasingly used. However, training such neural controllers is complicated by the absence of labeled datasets for controlled objects at the stage of creating the system. This work investigates model-oriented training of a neural controller and the creation of an initial training dataset based on a simulation model of the system. For this purpose, the problem of synthesizing optimal statistics of input data of the simulation model according to the criterion of minimum average losses is posed and solved. The problem of generating a dataset with specified statistical characteristics for training a controller of a dynamic control system is complicated by the fact that it is necessary to simultaneously provide both the necessary multidimensional probability distribution in the space of realizations of the vector influence on the system and its correlation structure in time. The mean square error of the neural controller is chosen as the loss function. Averaging is performed both over a set of realizations and over time to take into account the dynamics of the system. The simulation model of the decentralized system is created on the Scilab/Xcos platform using a pre-created library of blocks for modeling distributed systems. The task of generating a dataset with a given probability distribution and correlation structure is solved by generating an uncorrelated uniformly distributed dataset with subsequent nonlinear dynamic transformation. Simultaneously obtaining a given type of statistical distribution and correlation structure is obtained by iterative means with gradual approximation to the required result.