Functional interfaces for reference classes

Question

I am trying to figure out the conceptual implications of providing a functional user interface to reference classes (or indeed S4/S3 classes). In short, I am concerned that if I write code that looks like as below, then it is equivalent to writing functional code.

Here is a simple linear regression class:

linRegClass = setRefClass('linRegClass',
                          fields = list(formulaReg = 'formula',
                                        dataReg = 'data.frame'),
                          methods = list(doReg = function() {
                            lm(.self$formulaReg, data = .self$dataReg)
                          }))

linRegInstance = linRegClass$new(dataReg = cars, 
                    formulaReg = as.formula(speed ~ dist))
linRegInstance$doReg()
class(linRegInstance)

The object-oriented interface is not very user friendly, so as in Martin Morgan's slides, I will write a functional user interface for the underlying reference class:

fnLinReg = function(formulaReg, dataReg) {
  linRegInstance = linRegClass$new(formulaReg = formulaReg, 
                                   dataReg = dataReg)
  linRegInstance$doReg()
}

## use the functional interface
fnLinReg(dataReg = cars, formulaReg = as.formula(speed ~ dist))

Now this functional interface is observationally equivalent to a purely functional

fnLinReg2 = function(formulaReg, dataReg) {
  lm(formula = formulaReg, data = dataReg)
}

## use the pure function
fnLinReg2(dataReg = cars, formulaReg = as.formula(speed ~ dist))

I am trying to figure out whether this is because my example is pathologically simple, but I still wonder if there is any point in writing reference classes and then wrapping them in functional interfaces versus writing purely functional code.

Any detailed examples would help greatly.

Your example very stilted. Any kind of OO is bad for this example: you're not storing the class and using its methods later, you simply want the call to `lm`, so of course the pure function will be better. — Scott Ritchie, Dec 18 '13 at 06:45

Scott Ritchie · Accepted Answer · 2013-12-18T07:45:54.033

1

Complicated constructors can be defined using the initialize method, which will automatically be called when you call new. I've modified your example to contain the initialize method, and a new field to store the regression result, because initialize will always return the referenceClass:

linRegClass = setRefClass('linRegClass',
                          fields = list(formulaReg = 'formula',
                                        dataReg = 'data.frame',
                                        result = 'lm'),
                          methods = list(doReg = function() {
                            lm(.self$formulaReg, data = .self$dataReg)
                          },
                          initialize = function(formulaReg, dataReg, ...) {
                            formulaReg <<- formulaReg
                            dataReg <<- dataReg
                            result <<- .self$doReg()
                          }))

So now we simply have:

linRegInstance <- linRegClass$new(dataReg = cars, 
                                    formulaReg = as.formula(speed ~ dist))
linRegInstance$result

edited Dec 18 '13 at 07:45

answered Dec 18 '13 at 06:41

Scott Ritchie

10,293
3
28
64

less verbose variable names and class names also would help! – Scott Ritchie Dec 18 '13 at 06:42
Ha, I refuse to use less descriptive variable names. ;) Thanks for your answer. Could you give an example of a more elaborate example where interfacing to a reference class makes sense? – tchakravarty Dec 18 '13 at 06:50
It makes more sense to use it when you have an object, rather than a function. Like a particular data-type, or dataset (if you have a set of common analyses you want to perform). – Scott Ritchie Dec 18 '13 at 06:59
Ya, it is kinda dawning on me now. Thanks. – tchakravarty Dec 18 '13 at 06:59
1

By the way, you will run into [this issue](http://stackoverflow.com/q/18893290/1414455) with complex constructors with no defaults. – tchakravarty Dec 18 '13 at 07:14
There's two parts to that. One: `...` should be added as an extra argument to `initialize` for inheritance to work properly. I forgot that, so I've edited my answer. Two, in the linked question's case, they should really define an pseudo-abstract class that is parent to both. – Scott Ritchie Dec 18 '13 at 07:47
I actually have no idea if R can handle things like abstract classes or private methods. – Scott Ritchie Dec 18 '13 at 07:48
No private methods, that's for sure. – tchakravarty Dec 18 '13 at 07:51
I'm sure there's nothing stopping you from instantiating a class you've designed to be abstract either ;) – Scott Ritchie Dec 18 '13 at 10:03

Functional interfaces for reference classes

1 Answers1