Abstract:
A stochastic algorithm is proposed and analyzed, that is a probabilistic generalization of gradient method, for solving convex models. A random change of „old” partial derivatives with „new” ones is performed from one iteration to another. Convergence aspects of this scheme are analyzed for the case when the step is adjusted programmatically.
Certain conditions are indicated, that, being respected, ensure the convergence of this scheme to the optimal solution with probability 1.
