Application of Semi-Supervised Graph-Based Learning for the Classification of Solutions of Dynamical Systems
Keywords: Semi-Supervised Learning, Laplace Learning, Poisson Learning, Dynamical System
Abstract: In this article, we consider the application of one of the new methods of semi-supervised graph-based machine learning based (GSSL) on label propagation using the Poisson equation to solve the problem of classifying solutions of dynamic systems by two-dimensional Poincaré section data.
The used Poisson learning has advantages over classical Laplace learning since it allows using significantly less labeled data to achieve the desired accuracy of classification and to spent less time.
The article proposes a modification of Poisson learning and uses the Nesterov algorithm to find the minimum loss function, to improve the convergence results compared to the classical gradient descent and achieves an accuracy of 92\% even with 10 labels per class taking into account accumulation errors and time constraints. Which is acceptable for solving our general task - building an automated system for classifying solutions of dynamic systems in real time.
Submission Number: 51
Loading