The stochastic quasigradient projection method

From MaRDI portal





For the minimization of the convex regression function \(f(x)=Mf(x,\theta)\) on a convex compactum X from \(E^ n\), an iterative method has been presented by \textit{Yu. M. Ermol'ev} [Methods of stochastic programming (Russian) (1976; Zbl 0484.90070)], called the method of the projection of the stochastic quasigradient, which is the analogue of the method of the generalized gradient: \[ x^{s+1}=\pi_ X(x^ s-\rho_ s\xi^ s),\quad s=0,1,.... \] Here \(\pi_ X(y^ s)\) is the projection of the point \(y^ s\) onto X (the point of the wet X that is closest to \(y^ s)\). In this paper, for the minimization on a convex compactum of the regression function F(x), satisfying a local Lipschitz condition, we present a method which also makes use of the projection operation onto X. We note that the function F(x) need not be convex. Unlike the method given by Yu. M. Ermol'ev [loc. cit.] and its variant for convex locally Lipschitz functions [see the first author, in: Numerical Methods of Nonlinear Programming, Izd. Khar'kov. Gos. Univ., Khar'kov, 41-43 (Russian) (1979)], the algorithm presented here is based on the nonrelaxational analogue of the gradient projection method [see \textit{V. G. Karmanov}, Mathematical programming (Russian) (1980; Zbl 0496.90052)], known as Rosen's method. This method is called the stochastic quasigradient projection method. In this method, instead of the point \(y^ s=x^ s-\rho_ s\cdot \xi^ s\) one projects onto the set X the direction of \(\xi^ s\), the stochastic quasigradient.











This page was built for publication: The stochastic quasigradient projection method

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3321842)