A step size rule for unconstrained optimization

This short paper proposes an alternative rule for choosing the step size in the line search associated with unconstrained nonlinear minimization of \(f(x)\), \(x\in \mathbb{R}^ n\). Given a descent direction \(d_ i\), the usual line search phase is: find \(\alpha_ i\geq 0\) satisfying \(f(x_ i,\alpha_ id_ i)=min\{f(x_ i+\alpha d_ i), \alpha\geq 0\}\). In general, an inexact search has to be performed. The author's alternative line search introduces a parameter \(\beta_ j\), from which \(\gamma_ j=f(x_ i+\beta_ jd_ i)-f(x_ i)-\beta_ jd_ i^ Tg_ i\) is computed \((g_ i=\nabla f(x_ i))\) as part of the modified search procedure. It is proved that the resultant sequence \(\{f(x_ i)\}\) is decreasing and that, if \(f(x)\) is strictly convex quadratic, the alternative line search is exact.

Recommendations

Cites work

Cited in

(11)

This page was built for publication: A step size rule for unconstrained optimization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q1207052)