Састанак Одељења за математику, САНУ, 1. октобар 2009.

• 28. Септембар, 2009
• Коментари (0)

Andreas Griewank,  Humboltd University Berlin: "DERIVATIVE-BASED OPTIMIZATION"

Предавање ће се одржати у четврак, 1.10.2009.  у 14 часова, у сали 2 САНУ.

Abstract:Since Cauchy's time standard optimization methods require the evaluation of first derivatives for objective and constraint functions. At least the existence of these gradients and Jacobians istypically needed to ensure convergence for any optimization method that does not simply cover the feasible domain with sample points. The latter shot-gun approach is clearly unsuitable for high-dimensional problems of current interest.To beat the 'curse of dimensionality' in nonlinear optimization one can exploit the fact that continuous and discrete adjoints alike yield first and second derivative vectors at costs proportional to the underlying function and constraint evaluations. We discuss the resulting 'selective derivative methods' in comparison to derivative-free-methods in terms of numerical efficiency, applicability, and user convenience. Finally we report numerical experiences from aerodynamics and oceanography as depicted on the graph.