topic Re: Looking for fastest optimization subroutines in SAS/IML Software and Matrix Computations https://communities.sas.com/t5/SAS-IML-Software-and-Matrix/Looking-for-fastest-optimization-subroutines/m-p/92606#M602 <HTML><HEAD></HEAD><BODY><P>Gosh, speed depends on so many things, you should really time it for the problem you have with, say, 10% of the data.</P><P></P><P>Given that you have Jacobian and Hessian information, I'd probably use one of the Newton techniques. The Newton-Raphons (NLPNRA) algorithm has quadratic convergence near the solution, so I use it when I can get a good starting guess.&nbsp;&nbsp; I'd&nbsp; use&nbsp; the quasi-Newton algorithm (NLPQN) when the Hessian matrix is much more expensive to compute than the function and derivatives.</P></BODY></HTML> Sun, 02 Jun 2013 10:12:58 GMT Rick_SAS 2013-06-02T10:12:58Z Looking for fastest optimization subroutines https://communities.sas.com/t5/SAS-IML-Software-and-Matrix/Looking-for-fastest-optimization-subroutines/m-p/92605#M601 <HTML><HEAD></HEAD><BODY><P>I am performing an optimization with a large set of data.&nbsp; I can provide analytic gradient and hessian modules, but they are computationally intensive, like the function module.&nbsp; Of the optimization subroutines in IML (call NLP__), which one is the fastest?</P></BODY></HTML> Sat, 01 Jun 2013 22:53:51 GMT https://communities.sas.com/t5/SAS-IML-Software-and-Matrix/Looking-for-fastest-optimization-subroutines/m-p/92605#M601 opti_miser 2013-06-01T22:53:51Z Re: Looking for fastest optimization subroutines https://communities.sas.com/t5/SAS-IML-Software-and-Matrix/Looking-for-fastest-optimization-subroutines/m-p/92606#M602 <HTML><HEAD></HEAD><BODY><P>Gosh, speed depends on so many things, you should really time it for the problem you have with, say, 10% of the data.</P><P></P><P>Given that you have Jacobian and Hessian information, I'd probably use one of the Newton techniques. The Newton-Raphons (NLPNRA) algorithm has quadratic convergence near the solution, so I use it when I can get a good starting guess.&nbsp;&nbsp; I'd&nbsp; use&nbsp; the quasi-Newton algorithm (NLPQN) when the Hessian matrix is much more expensive to compute than the function and derivatives.</P></BODY></HTML> Sun, 02 Jun 2013 10:12:58 GMT https://communities.sas.com/t5/SAS-IML-Software-and-Matrix/Looking-for-fastest-optimization-subroutines/m-p/92606#M602 Rick_SAS 2013-06-02T10:12:58Z