Make `add_stable_points` optional in `IndShockConsumer`
Closed this issue · 8 comments
Mv77 commented
The solver for our most widely used class IndShockConsumerType calculates various points of theoretical interest via its method add_stable_points.
HARK/HARK/ConsumptionSaving/ConsIndShockModel.py
Line 1078 in 7ce7138
@mnwhite recently pointed out that this method can take up to 70% of the total time that it takes to solve a model.
These points are often not needed and, as user, I would like to have the option to not try to find them if I am doing structural estimation or calculating a Jacobian.
I want to propose that we disable the calculation of these points in the default solver, keeping them as an option to be activated by the user if he wants to calculate them.
Mv77 commented
Marking with high priority as this is taking up 70% of the computing time of our most popular agent type.
mnwhite commented
I have a branch where this is taken care of. It can be finished up and
merged in once the check_conditions branch is merged.
…On Thu, Aug 24, 2023, 3:08 AM Mateo Velásquez-Giraldo < ***@***.***> wrote:
Marking with high priority as this is taking up 70% of the computing time
of our most popular agent type.
—
Reply to this email directly, view it on GitHub
<#1344 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ADKRAFIE32U6GPO4WHZ4A6DXW2SSXANCNFSM6AAAAAA34IFFWE>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
llorracc commented
If someone is solving an infinite horizon problem, let's have a two step
process:
1. Quickly-checkable convergence checked by the existing distance metric,
which has no particular justification as being meaningful (because for
example it gives equal weight to aGrid points THAT WILL NEVER EVER BE
REACHED to grid points that ARE REACHED ALL THE TIME).
2. If it converges according to the cheap metric, we test whether it
converges according to the better but slower metric
- If so, we're done
- if not, tighten the cheap metric by a factor of ten and
resume convergence.
This more sophisticated thing should be our default for problems that
satisfy the GIC-Mod
On Thu, Aug 24, 2023 at 7:42 AM Matthew N. White ***@***.***>
wrote:
… I have a branch where this is taken care of. It can be finished up and
merged in once the check_conditions branch is merged.
On Thu, Aug 24, 2023, 3:08 AM Mateo Velásquez-Giraldo <
***@***.***> wrote:
> Marking with high priority as this is taking up 70% of the computing
time
> of our most popular agent type.
>
> —
> Reply to this email directly, view it on GitHub
> <#1344 (comment)>,
> or unsubscribe
> <
https://github.com/notifications/unsubscribe-auth/ADKRAFIE32U6GPO4WHZ4A6DXW2SSXANCNFSM6AAAAAA34IFFWE>
> .
> You are receiving this because you were mentioned.Message ID:
> ***@***.***>
>
—
Reply to this email directly, view it on GitHub
<#1344 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAKCK74UDDEMXJU5UCEBBKTXW3SULANCNFSM6AAAAAA34IFFWE>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
--
- Chris Carroll
mnwhite commented
That requires a significant extension to how our convergence rule works,
but it can be done in principle.
On Thu, Aug 24, 2023 at 11:17 AM Christopher Llorracc Carroll <
***@***.***> wrote:
… If someone is solving an infinite horizon problem, let's have a two step
process:
1. Quickly-checkable convergence checked by the existing distance metric,
which has no particular justification as being meaningful (because for
example it gives equal weight to aGrid points THAT WILL NEVER EVER BE
REACHED to grid points that ARE REACHED ALL THE TIME).
2. If it converges according to the cheap metric, we test whether it
converges according to the better but slower metric
- If so, we're done
- if not, tighten the cheap metric by a factor of ten and
resume convergence.
This more sophisticated thing should be our default for problems that
satisfy the GIC-Mod
On Thu, Aug 24, 2023 at 7:42 AM Matthew N. White ***@***.***>
wrote:
> I have a branch where this is taken care of. It can be finished up and
> merged in once the check_conditions branch is merged.
>
> On Thu, Aug 24, 2023, 3:08 AM Mateo Velásquez-Giraldo <
> ***@***.***> wrote:
>
> > Marking with high priority as this is taking up 70% of the computing
> time
> > of our most popular agent type.
> >
> > —
> > Reply to this email directly, view it on GitHub
> > <#1344 (comment)>,
> > or unsubscribe
> > <
>
https://github.com/notifications/unsubscribe-auth/ADKRAFIE32U6GPO4WHZ4A6DXW2SSXANCNFSM6AAAAAA34IFFWE>
>
> > .
> > You are receiving this because you were mentioned.Message ID:
> > ***@***.***>
> >
>
> —
> Reply to this email directly, view it on GitHub
> <#1344 (comment)>,
> or unsubscribe
> <
https://github.com/notifications/unsubscribe-auth/AAKCK74UDDEMXJU5UCEBBKTXW3SULANCNFSM6AAAAAA34IFFWE>
> .
> You are receiving this because you are subscribed to this thread.Message
> ID: ***@***.***>
>
--
- Chris Carroll
—
Reply to this email directly, view it on GitHub
<#1344 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ADKRAFPDCLGYDOYISZ4EFZLXW4LY7ANCNFSM6AAAAAA34IFFWE>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
Mv77 commented
An issue is that the stable points are being searched for even in life cycle calibrations.
mnwhite commented
Right. I would prefer to move that code *entirely* out of the solver and
have it be handled at the AgentType level, as something that can be run
post-convergence. With a broader framework for convergence, additional
conditions can be checked after it converges in a cFunc sense.
…On Thu, Aug 24, 2023, 12:27 PM Mateo Velásquez-Giraldo < ***@***.***> wrote:
An issue is that the stable points are being searched for even in life
cycle calibrations.
—
Reply to this email directly, view it on GitHub
<#1344 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ADKRAFI62DVUH546Y2JSJTTXW4UBVANCNFSM6AAAAAA34IFFWE>
.
You are receiving this because you were mentioned.Message ID:
***@***.***>
llorracc commented
It should be done.
On Thu, Aug 24, 2023 at 11:50 AM Matthew N. White ***@***.***>
wrote:
… That requires a significant extension to how our convergence rule works,
but it can be done in principle.
On Thu, Aug 24, 2023 at 11:17 AM Christopher Llorracc Carroll <
***@***.***> wrote:
> If someone is solving an infinite horizon problem, let's have a two step
> process:
>
> 1. Quickly-checkable convergence checked by the existing distance
metric,
> which has no particular justification as being meaningful (because for
> example it gives equal weight to aGrid points THAT WILL NEVER EVER BE
> REACHED to grid points that ARE REACHED ALL THE TIME).
> 2. If it converges according to the cheap metric, we test whether it
> converges according to the better but slower metric
> - If so, we're done
> - if not, tighten the cheap metric by a factor of ten and
> resume convergence.
>
> This more sophisticated thing should be our default for problems that
> satisfy the GIC-Mod
>
>
>
> On Thu, Aug 24, 2023 at 7:42 AM Matthew N. White ***@***.***>
> wrote:
>
> > I have a branch where this is taken care of. It can be finished up and
> > merged in once the check_conditions branch is merged.
> >
> > On Thu, Aug 24, 2023, 3:08 AM Mateo Velásquez-Giraldo <
> > ***@***.***> wrote:
> >
> > > Marking with high priority as this is taking up 70% of the computing
> > time
> > > of our most popular agent type.
> > >
> > > —
> > > Reply to this email directly, view it on GitHub
> > > <
#1344 (comment)>,
>
> > > or unsubscribe
> > > <
> >
>
https://github.com/notifications/unsubscribe-auth/ADKRAFIE32U6GPO4WHZ4A6DXW2SSXANCNFSM6AAAAAA34IFFWE>
>
> >
> > > .
> > > You are receiving this because you were mentioned.Message ID:
> > > ***@***.***>
> > >
> >
> > —
> > Reply to this email directly, view it on GitHub
> > <#1344 (comment)>,
> > or unsubscribe
> > <
>
https://github.com/notifications/unsubscribe-auth/AAKCK74UDDEMXJU5UCEBBKTXW3SULANCNFSM6AAAAAA34IFFWE>
>
> > .
> > You are receiving this because you are subscribed to this
thread.Message
> > ID: ***@***.***>
> >
>
>
> --
> - Chris Carroll
>
> —
> Reply to this email directly, view it on GitHub
> <#1344 (comment)>,
> or unsubscribe
> <
https://github.com/notifications/unsubscribe-auth/ADKRAFPDCLGYDOYISZ4EFZLXW4LY7ANCNFSM6AAAAAA34IFFWE>
> .
> You are receiving this because you were mentioned.Message ID:
> ***@***.***>
>
—
Reply to this email directly, view it on GitHub
<#1344 (comment)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AAKCK77N25NQSAGCLWECLVTXW4PU7ANCNFSM6AAAAAA34IFFWE>
.
You are receiving this because you commented.Message ID:
***@***.***>
--
- Chris Carroll
mnwhite commented
This was taken care of a while back, closing.