The study numerical to detect collisions using neural network with D* algorithm

Abstract

A numerical study of the proposed iterative algorithm using the Matlab complex is carried out. During the study, the task of teaching a neural network to plan a route was solved in the same way as the "teacher" algorithm, for which D* was chosen. The initial filtering option is selected in such a way that, when a collision occurs, a trajectory point that preceded the collision is fed into the filtered sample, i.e., in which the neural network incorrectly classified the situation. This method of filtering was not effective, because the wrong decision that led to the collision could have been made not immediately before the collision, but earlier. In this regard, the procedure for filtering examples for training has been modified so that when a collision occurs, maps with the robot's position at all points of its trajectory are added to the training sample. This allows you to significantly increase the success of achieving the goal.