New samplers and mode compute methods

Estimation ML/Bayesian estimation
#1

Dear All,
I was reading slides on non-linear estimation from the 2019 Dynare Summer-school and I saw you were already testing a version of Herbst and Shorfheide SMC sampler based on tempered likelihood and a Dynamic Striated MH sampler.
Are they available in Dynare 4.6.3?
In positive case, what are the settings to call them?

Also, I saw in the 4.6 manual, there is a new non-gradient based algorithm for computing the mode using the particle-swarm Matlab optimization algorithms (mode_compute = 12).
When is this method more advised than others?

Many thanks in advance for your availability!

#2
  1. From what I can see, the codes are there, e.g. https://git.dynare.org/Dynare/particles/-/commit/22be3797a87b2f9ce5ae75556ab18774088d66ee
    However, there is no documentation and no examples yet. So getting them to work may be challenging. Maybe @frederic.karame can help.
  2. The particle swarm is a global optimizer. So you can give it a try in challenging situations. But my experience is that it tends to be really slow.
#3

Herbst and Shorfheide SMC sampler and the Dynamic Striated MH sampler have only been tested on the toy model I use at the summer school, for a few filters. But we also need to develop an interface to use them and to add the documentation…

#4

@frederic.karame Could you provide the toy model files?

#5

@jpfeifer The toy model is dsge_base2.mod.

#6

I meant how to invoke the two samplers. https://git.dynare.org/Dynare/dynare/-/blob/master/tests/particle/dsge_base2.mod does not seem to have those cases included.

#7

To test, I put a call in dynare_estimation1.m

Herbst_Schorfheide_sampler(TargetFun,xparam1,mh_bounds,dataset_,dataset_info,options_,M_,estim_params_,bayestopt_,oo_)
or
DSMH_sampler(TargetFun,xparam1,mh_bounds,dataset_,dataset_info,options_,M_,estim_params_,bayestopt_,oo_)

a priori everything is defined in Global_initialization.m with comprehensive notations (DSMH_ for DSMH) and HB_ for Herbst &Schorfheide