Yakubovich's method of recursive objective inequalities in machine learning
Mikhail Lipkovich · IFAC-PapersOnLine · 2022
“Stripe” algorithm was designed by V.A. Yakubovich as an algorithm for linear classification model training. The basic approach is to reduce loss minimization problem to solving a system of infinite number of inequalities. The present paper considers various forms of this reduction as well as its practical applicability. Obtained algorithms are experimentally compared with traditional linear models such as logistic regression and linear regression trained using stochastic gradient descent. Simulation results show that the “Stripe” algorithm possesses fast convergence and in particular suitable well in the paradigm of online machine learning.