Resumo: This talk addresses unconstrained minimization and nonlinear least-squares problems, assuming Lipschitz continuity of the Hessian or the Jacobian, respectively. Regularizing rules are devised so that the full Newton-like step or the full Levenberg-Marquardt step are likely to be accepted by the Armijo sufficient descent condition. Indeed, under availability of the Lipschitz constant, full steps are accepted. Otherwise, an estimate of such a constant is dynamically updated. Convergence properties are analyzed and the numerical performance is investigated, with encouraging results.
Palestrante: Elizabeth Wegner Karas, Departamento de Matemática, UFPR
Local: Auditório (LAED) do Departamento de Matemática – Térreo