8:03 p.m.
Hi,
I had a problem with a mle estimation today. My script canceled execution with the
following error message:
Using numerical derivatives
Iteration 1: Log-likelihood = 7274,73993927
Parameters: 3,4478 -0,17644 6,7838
Gradients: 2,6526e+06 2,6526e+06 -2,6526e+06
failed to match initial value of objective function, 7274,74
--- FINAL VALUES:
Log-likelihood = 7274,73993926 (steplength = 1,04858e-14)
Parameters: 3,4478 -0,17644 6,7838
Gradients: 2,6526e+06 2,6526e+06 -2,6526e+06
Tolerance = 1,81899e-12
The convergence criterion was not met
seems like the starting set was already quite optimal. Though gretl steps to the new
point, because the criteria for advancing in line 544 in gretl_bfgs.c is (f >= fmax +
d). d is set as d = sumgrad * steplen * acctol and in my case seems to be negative. The
criteria for printing the error message (line 645) is: (fmax < f0). The attached patch
(one-liner) changes this to ((fmax + d) <= f0). Is this change ok?
Christoph
8:43 p.m.
On Thu, 2 Jul 2009, Klein, Christoph wrote:
I had a problem with a mle estimation today. My script canceled
execution with the following error message:
Using numerical derivatives
Iteration 1: Log-likelihood = 7274,73993927
Parameters: 3,4478 -0,17644 6,7838
Gradients: 2,6526e+06 2,6526e+06 -2,6526e+06
failed to match initial value of objective function, 7274,74
--- FINAL VALUES:
Log-likelihood = 7274,73993926 (steplength = 1,04858e-14)
Parameters: 3,4478 -0,17644 6,7838
Gradients: 2,6526e+06 2,6526e+06 -2,6526e+06
Tolerance = 1,81899e-12
The convergence criterion was not met
seems like the starting set was already quite optimal. Though
gretl steps to the new point, because the criteria for advancing
in line 544 in gretl_bfgs.c is (f >= fmax + d). d is set as d =
sumgrad * steplen * acctol and in my case seems to be negative.
The criteria for printing the error message (line 645) is: (fmax
< f0). The attached patch (one-liner) changes this to ((fmax +
d) <= f0). Is this change ok?
I think you're right. I've applied the patch -- thanks!
Allin
10:13 p.m.
New subject: AW: bfgs: failed to match initial value
i'm sorry but my patch doesn't solve the problem,possibly making the problem
worse. I will look tomorrow at the problem again.
Can you revert the change, please?!
Christoph
________________________________________
Von: gretl-devel-bounces(a)lists.wfu.edu [gretl-devel-bounces(a)lists.wfu.edu] im Auftrag von
Allin Cottrell [cottrell(a)wfu.edu]
Gesendet: Donnerstag, 2. Juli 2009 22:43
An: Gretl development
Betreff: Re: [Gretl-devel] bfgs: failed to match initial value
On Thu, 2 Jul 2009, Klein, Christoph wrote:
I had a problem with a mle estimation today. My script canceled
execution with the following error message:
Using numerical derivatives
Iteration 1: Log-likelihood = 7274,73993927
Parameters: 3,4478 -0,17644 6,7838
Gradients: 2,6526e+06 2,6526e+06 -2,6526e+06
failed to match initial value of objective function, 7274,74
--- FINAL VALUES:
Log-likelihood = 7274,73993926 (steplength = 1,04858e-14)
Parameters: 3,4478 -0,17644 6,7838
Gradients: 2,6526e+06 2,6526e+06 -2,6526e+06
Tolerance = 1,81899e-12
The convergence criterion was not met
seems like the starting set was already quite optimal. Though
gretl steps to the new point, because the criteria for advancing
in line 544 in gretl_bfgs.c is (f >= fmax + d). d is set as d =
sumgrad * steplen * acctol and in my case seems to be negative.
The criteria for printing the error message (line 645) is: (fmax
< f0). The attached patch (one-liner) changes this to ((fmax +
d) <= f0). Is this change ok?
I think you're right. I've applied the patch -- thanks!
Allin
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
11:56 p.m.
New subject: AW: bfgs: failed to match initial value
On Fri, 3 Jul 2009, Klein, Christoph wrote:
i'm sorry but my patch doesn't solve the problem,possibly
making
the problem worse. I will look tomorrow at the problem again.
Can you revert the change, please?!
OK, done. I'll try taking a closer look myself. As the comment
in the code suggests -- "FIXME this should never happen" -- the
issue we're looking at is one that rarely arises, and is somewhat
puzzling.
Allin.
12:11 a.m.
New subject: AW: bfgs: failed to match initial value
On Thu, 2 Jul 2009, Allin Cottrell wrote:
On Fri, 3 Jul 2009, Klein, Christoph wrote:
> i'm sorry but my patch doesn't solve the problem,possibly making
> the problem worse. I will look tomorrow at the problem again.
> Can you revert the change, please?!
OK, done. I'll try taking a closer look myself. As the comment
in the code suggests -- "FIXME this should never happen" -- the
issue we're looking at is one that rarely arises, and is somewhat
puzzling.
Let me be a bit more explicit. The reason why this is puzzling is
that, typically, if you take MLE results and pass them back as
initial values for the same problem, gretl's BFGS reports
immediate convergence, not the error you're talking about ("failed
to match initial value of objective function"). For example, this
simple script:
open data9-7
arma 1 1 ; QNC
matrix b = $coeff
set initvals b
arma 1 1 ; QNC -v
Jack, can you help us out on this?
Allin.
7:35 a.m.
New subject: AW: bfgs: failed to match initial value
On Thu, 2 Jul 2009, Allin Cottrell wrote:
On Thu, 2 Jul 2009, Allin Cottrell wrote:
> On Fri, 3 Jul 2009, Klein, Christoph wrote:
>
>> i'm sorry but my patch doesn't solve the problem,possibly making
>> the problem worse. I will look tomorrow at the problem again.
>> Can you revert the change, please?!
>
> OK, done. I'll try taking a closer look myself. As the comment
> in the code suggests -- "FIXME this should never happen" -- the
> issue we're looking at is one that rarely arises, and is somewhat
> puzzling.
Let me be a bit more explicit. The reason why this is puzzling is
that, typically, if you take MLE results and pass them back as
initial values for the same problem, gretl's BFGS reports
immediate convergence, not the error you're talking about ("failed
to match initial value of objective function"). For example, this
simple script:
open data9-7
arma 1 1 ; QNC
matrix b = $coeff
set initvals b
arma 1 1 ; QNC -v
Jack, can you help us out on this?
It would reaaly help to have a testcase, to see exactly what's going on
here. Christoph, I noticed in your original post that the gradients look
funny:
- they are big in absolute value, which shouldn't happen if the starting
point is optimal
- they look the same (again, in absolute value), which is typically what
happens when your model is underidentified, but not necessarily.
Also, IIRC you're using numerical derivatives: there may be a problem with
those too, perhaps.
So if you could post a testcase, it would be most helpful.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
8:35 p.m.
New subject: AW: AW: bfgs: failed to match initial value
I'm sorry for all the trouble. Looking carefully through my code, it seems more like
the problem is within my scripts than in gretl. I will try to find out why this problem
did arise in the next days.
Sorry again and thanks for your helping pointers!
Christoph
________________________________________
Von: gretl-devel-bounces(a)lists.wfu.edu [gretl-devel-bounces(a)lists.wfu.edu] im Auftrag von
Riccardo (Jack) Lucchetti [r.lucchetti(a)univpm.it]
Gesendet: Freitag, 3. Juli 2009 09:35
An: Gretl development
Betreff: Re: AW: [Gretl-devel] bfgs: failed to match initial value
On Thu, 2 Jul 2009, Allin Cottrell wrote:
On Thu, 2 Jul 2009, Allin Cottrell wrote:
> On Fri, 3 Jul 2009, Klein, Christoph wrote:
>
>> i'm sorry but my patch doesn't solve the problem,possibly making
>> the problem worse. I will look tomorrow at the problem again.
>> Can you revert the change, please?!
>
> OK, done. I'll try taking a closer look myself. As the comment
> in the code suggests -- "FIXME this should never happen" -- the
> issue we're looking at is one that rarely arises, and is somewhat
> puzzling.
Let me be a bit more explicit. The reason why this is puzzling is
that, typically, if you take MLE results and pass them back as
initial values for the same problem, gretl's BFGS reports
immediate convergence, not the error you're talking about ("failed
to match initial value of objective function"). For example, this
simple script:
open data9-7
arma 1 1 ; QNC
matrix b = $coeff
set initvals b
arma 1 1 ; QNC -v
Jack, can you help us out on this?
It would reaaly help to have a testcase, to see exactly what's going on
here. Christoph, I noticed in your original post that the gradients look
funny:
- they are big in absolute value, which shouldn't happen if the starting
point is optimal
- they look the same (again, in absolute value), which is typically what
happens when your model is underidentified, but not necessarily.
Also, IIRC you're using numerical derivatives: there may be a problem with
those too, perhaps.
So if you could post a testcase, it would be most helpful.
Riccardo (Jack) Lucchetti
Dipartimento di Economia
Università Politecnica delle Marche
r.lucchetti(a)univpm.it
http://www.econ.univpm.it/lucchetti
9:22 a.m.
New subject: AW: AW: bfgs: failed to match initial value
Am 03.07.2009 22:35, Klein, Christoph schrieb:
I'm sorry for all the trouble. Looking carefully through my code,
it seems more like the problem is within my scripts than in gretl. I will try to find out
why this problem did arise in the next days.
Sorry again and thanks for your helping pointers!
Christoph
Nevertheless, if it was the case that you get an error message "which
should never happen" then it's worthwhile to fix it, no? Or am I mixing
things up?
-sven
7:21 p.m.
New subject: AW: AW: bfgs: failed to match initial value
On Sat, 4 Jul 2009, Sven Schreiber wrote:
Am 03.07.2009 22:35, Klein, Christoph schrieb:
> I'm sorry for all the trouble. Looking carefully through my code, it seems more
like the problem is within my scripts than in gretl. I will try to find out why this
problem did arise in the next days.
>
> Sorry again and thanks for your helping pointers!
>
> Christoph
Nevertheless, if it was the case that you get an error message "which
should never happen" then it's worthwhile to fix it, no?
Yes, I think it would still be useful to have a testcase which
provokes that error message.
Allin.
6:19 a.m.
Am 02.07.2009 22:03, Klein, Christoph schrieb:
seems like the starting set was already quite optimal. Though gretl
steps to the new point, because the criteria for advancing in line
544 in gretl_bfgs.c is (f >= fmax + d). d is set as d = sumgrad *
steplen * acctol and in my case seems to be negative. The criteria
for printing the error message (line 645) is: (fmax < f0). The
attached patch (one-liner) changes this to ((fmax + d) <= f0). Is
this change ok?
Without really knowing the code context and thus at the risk of babbling
stupid things, shouldn't the case of d<0 then be ruled out? That is,
change the criterion in line 544 to f >= fmax + max(0,d)?
Because otherwise it seems that starting from an optimum the negative
gradient moves you away from the maximum but f >= fmax + d will be true
when d<0 and thus gretl thinks it's going in the right direction.
just a thought,
sven
6:34 a.m.
New subject: AW: bfgs: failed to match initial value
The applied patch alters the criteria for issuing an "failed to macht initital
vaue"-error again. d is only subtracted from fmax, if d was used, i.e. crit_ok is
true.
But if Sven is right, the patch doesn't solve the real issue here.
Christoph
________________________________________
Von: gretl-devel-bounces(a)lists.wfu.edu [gretl-devel-bounces(a)lists.wfu.edu] im Auftrag von
Sven Schreiber [svetosch(a)gmx.net]
Gesendet: Freitag, 3. Juli 2009 08:19
An: Gretl development
Betreff: Re: [Gretl-devel] bfgs: failed to match initial value
Am 02.07.2009 22:03, Klein, Christoph schrieb:
seems like the starting set was already quite optimal. Though gretl
steps to the new point, because the criteria for advancing in line
544 in gretl_bfgs.c is (f >= fmax + d). d is set as d = sumgrad *
steplen * acctol and in my case seems to be negative. The criteria
for printing the error message (line 645) is: (fmax < f0). The
attached patch (one-liner) changes this to ((fmax + d) <= f0). Is
this change ok?
Without really knowing the code context and thus at the risk of babbling
stupid things, shouldn't the case of d<0 then be ruled out? That is,
change the criterion in line 544 to f >= fmax + max(0,d)?
Because otherwise it seems that starting from an optimum the negative
gradient moves you away from the maximum but f >= fmax + d will be true
when d<0 and thus gretl thinks it's going in the right direction.
just a thought,
sven
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
7:58 a.m.
New subject: AW: bfgs: failed to match initial value
i think d is always negative, because sumgrad must be negative, because otherwise we
wouldn't pass line 526. steplen and acctol are always positive.
________________________________________
Von: gretl-devel-bounces(a)lists.wfu.edu [gretl-devel-bounces(a)lists.wfu.edu] im Auftrag von
Sven Schreiber [svetosch(a)gmx.net]
Gesendet: Freitag, 3. Juli 2009 08:19
An: Gretl development
Betreff: Re: [Gretl-devel] bfgs: failed to match initial value
Am 02.07.2009 22:03, Klein, Christoph schrieb:
seems like the starting set was already quite optimal. Though gretl
steps to the new point, because the criteria for advancing in line
544 in gretl_bfgs.c is (f >= fmax + d). d is set as d = sumgrad *
steplen * acctol and in my case seems to be negative. The criteria
for printing the error message (line 645) is: (fmax < f0). The
attached patch (one-liner) changes this to ((fmax + d) <= f0). Is
this change ok?
Without really knowing the code context and thus at the risk of babbling
stupid things, shouldn't the case of d<0 then be ruled out? That is,
change the criterion in line 544 to f >= fmax + max(0,d)?
Because otherwise it seems that starting from an optimum the negative
gradient moves you away from the maximum but f >= fmax + d will be true
when d<0 and thus gretl thinks it's going in the right direction.
just a thought,
sven
_______________________________________________
Gretl-devel mailing list
Gretl-devel(a)lists.wfu.edu
http://lists.wfu.edu/mailman/listinfo/gretl-devel
5649
days inactive
5651
days old
11 comments
4 participants
participants (4)
-
Allin Cottrell -
Klein, Christoph -
Riccardo (Jack) Lucchetti -
Sven Schreiber