Is there a Hessian-free optimization algorithm?

7 vues (au cours des 30 derniers jours)
Shahab
Shahab le 24 Mai 2016
I need to do some constrained optimization work on a function which is actually a Simulink model involving hundreds of variables. Each function evaluation takes about 6 or 7 seconds. So, lowering the number of function evaluations is critical to me.
I've looked into the available algorithms with fmincon (interior-point, sqp, active-set, trust-region-reflective) and they all require Hessian calculation.
Is there any algorithm available in MATLAB that can optimize the function via only function or gradient evaluation? (and does not require the time-consuming Hessian calculation)

Réponse acceptée

Sean de Wolski
Sean de Wolski le 24 Mai 2016
Modifié(e) : Sean de Wolski le 24 Mai 2016
In the Global Optimization Toolbox:
doc patternsearch
perhaps

Plus de réponses (1)

Mary Fenelon
Mary Fenelon le 27 Mai 2016
If your functions are smooth, a gradient-based solver such as fmincon may take fewer function evaluation overall. Some of its algorithms do not calculate the Hessian using finite differences but use an approximation to the Hessian; see more here . You might also want to review the docs on choosing an appropriate step size when optimizing a simulation to help decide between a gradient-based solver and patternsearch but it's hard to predict so if possible give them both a try.

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by