Centralizing Value and MargValue functions

[Mv77]

The trick of working with value and marginal value functions in 'inverse-space' is now applied in many models in HARK. This happens mostly through the ValueFunc and MargValueFunc classes.

However, the general idea of these classes is not captured in a centralized place. Different versions have been defined as models have been added: the 1-d version is in ConsIndShockModel, the 2-d version in ConsGenIncProcessModel, and my new portfolio model will use a 3-d version.

This makes it so that if I'm working on a new, say, 2-d model, I have to go check whether this has already been implemented in some of the existing models. Or if I need both 1-d and 2-d versions (as ConsPortfolioModel does), I need to import them from different models. I think this will lead to multiple versions of these objects being defined in different places.

This pr aims to construct general n-dimensional versions that can be imported from interpolation (or somewhere else?).

Would this be useful or is there a reason for these objects to be defined in different places?

[Mv77]

The previous commits centralize ValueFunc. All of its uses now point to a single version in interpolation.py.

If you like this idea, I can continue with MargValueFunc, which is trickier as some instances of its use need its derivatives. Still, I think the task is possible and worthy.

[sbenthall]

In my view this is exactly the right direction.

[Mv77]

The previous commits implement n-dimensional MargValueFunc and MargMargValueFunc, and replace their previous uses and definitions.

[sbenthall]

Would you like me to review this? I see it is not marked as WIP

[Mv77]

Yes, please

[sbenthall]

Beautiful work. Thanks for doing this. It passes review.

Could I ask you to please add a commit with an edit to the CHANGELOG describing the change and linking to this PR?

https://github.com/econ-ark/HARK/blob/master/Documentation/CHANGELOG.md

I believe it counts as a Major change. It should make it into the next release, which may be early this week.

[sbenthall]

One other question: I've noticed this related code in the PerfectForesight model which is involved in constructing value functions.

HARK/HARK/ConsumptionSaving/ConsIndShockModel.py

Line 438 in b45cabd

def defValueFuncs(self):

Is it possible to streamline that code any further given the changes you are making here?

[Mv77]

Is it possible to streamline that code any further given the changes you are making here?

Does not seem to me like it is. There are four lines of code, two using ValueFunc and MargValueFunc (that now will come from interpolation.py), the other two just defining a pair of points that are needed for ValueFunc in the special perfect foresight case.

[Mv77]

Beautiful work. Thanks for doing this. It passes review.

Could I ask you to please add a commit with an edit to the CHANGELOG describing the change and linking to this PR?

https://github.com/econ-ark/HARK/blob/master/Documentation/CHANGELOG.md

I believe it counts as a Major change. It should make it into the next release, which may be early this week.

Ready! (in the last commit)

[Mv77]

Call them ___CRRA.

[Mv77]

Since @sbenthall's review, I have:

• Added the change to the changelog.
• Changed the name of the functions to [az]*ValueFuncCRRA, to make it clear that CRRA utility is hardcoded in them.