Hi Michel,
not yet profiled the routines. With small models like ls2003 etc. is it quite competitive with numerical ones.
Also the actual dynare parallelization with the 'always-open' mode could be tried (also for numerical derivs).
best Marco
On 10/4/2011 3:40 PM, Michel Juillard wrote:
Thanks Marco,
this is maybe an area where we can thing about low level parallelization. Do you have already timing results? I always wondered about the gain in speed as the computation of the analytical derivatives is rather complex.
Best
Michel
On 10/4/2011 3:36 PM, Marco Ratto wrote:
Hi Michel,
by analytic Hessian I mean the matrix of second derivatives of the likelihood. the analytic asymptotic hessian is the limit of the hessian for infinite number of observations, where all second order derivative terms drop out in the formulas, and only first order derivatives are used to compute asymptotic hessian.
In the routines, for every time step of the kalman filter, I do a loop for every parameter value to compute the analytic score. If the hessian is desired, I also add the second loop for all cross derivatives.
In the end this is a loop similar to Harvey, since there is not true vectorization, but all common computations for the filter are done only once, instead of doing them many times for each derivative. So, computationally not a bad solution. I have to think if I can really vectorize the formulas and implement them as tensor/algebraic equations, but I am not sure if this really possible.
In the end, with one single call of kalman filter with the analytic scores/ hessian, I get the full result.
best Marco
On 10/4/2011 3:28 PM, Michel Juillard wrote:
Hi Marco,
what exactly do you call the analytic hessian and the analytic asymptotic hessian. I seem to remember from Harvey that you need to run the Kalman filter for each derivative of the likelihood with respect to one parameter. Could you vectorize these operations?
Best
Michel
On 10/4/2011 3:23 PM, Marco Ratto wrote:
Hi Stephane, Thanks a lot. By the way, last week in Lisbon I hve completed the routines to compute the analytic hessian (not only the analytic asymptotic hessian). If the likelihood functions are now fully updated, I can try to plug-in the analytic scores in optimization, if desired by the user. For the moment just for testing, with a syntax in the end. Moreover, I can allow for analytic HEssian in place of numerical one.
OK? best Marco
On 10/3/2011 12:37 PM, Stéphane Adjemian wrote:
Hi Marco,
I have just pushed a fix for your optimization routines (mode_compute==5) and added in the testsuite a mod file (fs2000d.mod) calling your routines.
Best, Stéphane.
Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev