New Approaches to the Approximation of Solutions in Machine Learning

Machine learning tasks focused on determining the laws of control of robots with complex locomotion are considered. The exponential computational complexity of such tasks is shown when using existing methods, in particular, reinforcement learning. The theoretical possibility of finding a multidimensional control function that is based on differential algebraic equations of the dynamics of such systems is substantiated by varying the selected subset of the constraints equations. The possibility of a significant reduction in the dimension of the parameter space of the optimization problem on this basis is analyzed. Examples of the use of the proposed method for solving problems of the dynamics of machines, zoomorphic and anthropomorphic robots are given. The universality of the method is shown in the sense of applicability to various technical systems, the equations of state of which are expressed in the form of differential algebraic equations. The method can be technically implemented in microprocessor control systems using neural networks trained on the proposed models.