Marco,
thanks for looking into the issue. Some of my changes are needed to get
rid of the global objective_function_penalty_base and pass it as an
argument.
Marco Ratto writes:
> Dear Michel,
>
> I think the code is OK without your changes, but indeed I changed
> the behaviour of the code w.r.t. stable version.
>
> Let me try to explain:
>
>
>
> 1) indeed the 6th input argument in newrat triggers how the hessian
> is computed in the final step.
>
> However, I realized that this is not the optimal way to go, so in
> dynare_minimize_objective this is better forced to be zero. The
> point is that with argument 0 (or 1) the best way to compute outer
> product gradient is with univarite kalman filter, since it
> maximizes the number of density points used in outer product.
I don't understand the previous point about univariate Kalman filter.
> However, when as usual multivariate filter is used for estimation, this would imply to change kalman_algo from within optimizer which is not a good thing to do. So, it is best to force zero in optimization. This implies in practice no extra computation in newrat.
> Then, in dynare_estimation_1.m, the need for that piece of code you mention at point 2).
> If newratflag = 0 and univariate filter was already used, nothing is done, and no further hessian computation is performed.
> If multivariate filter was used, then outer product gradient is re-done using univariate algorithm.
> Then, it seems also OK for me to trigger there full hessian computation with mr_hessian using newratflag=2 [although I must say I never do that].
>
>
>
>
> 2) as discussed in point 1), newratflag is indeed an option that triggers how the hessian in
> newrat is computed in place of the standard dynare one, but it needs
> to be used in
>
>
>
> ./matlab/dynare_estimation_1. lines 262-288
>
> and not in
>
>
> ./matlab/optimization/dynare_minimize_objective.m
>
> essentially linked to univariate kalman filter use in outer product gradient
>
> 3) I think re-setting analytic_derivative can be useful for subsequent
> computation, in order not to change options_ w.r.t. the user had set
> in the estimation() command.
When do we use analytic_derivative without the user indicating so in the
options?
Let's talk about it in Brussels.
Best
Michel
>
> thanks a lot
> best
> Marco
--
Michel Juillard