I love R, but it can be frustrating to code with and learn. Some of its datatypes are immensely powerful but work in mysterious ways. It allows you to manipulate expressions in a LISP macro-like fashion which, when used badly by library authors, can make many things appear magical (and inconsistant). There are many inconsistencies in the standard library because of different programming paradigms used (for example (s|m|t|r)apply() vs. Map() and filtering with df$col[] vs. subset() vs. Filter() ).

Yet I love writing code in it. So much can be done in so little code. I am always amazed at how little code I write to accomplish a task.

The RStudio IDE ( http://rstudio.org/ ) is a very pleasant environment to write code in.

Map and subset would be discouraged, they are just wrappers around the apply functions and [] anyway.

I wouldn't discourage them uniquivocally - for instance, 'subset' is an idiom often used in the context of data/relational tables. That matters a little bit (enough to pay a little penalty in raw performance), I would think.

What sort of tasks have you found where it really shines?

In some ways this is better, and in others it is much worse. Using named constants instead of magic numbers, and breaking up complex logic into simpler named parts is a huge win for readability/maintainability, as any programmer quickly learns. The big one liner would be a lot nicer in about 3-4 chunks.

I’d rewrite this example as something like:

  num.steps <- 1000
  num.walks <- 100
  step.std.dev <- 0.03
  start.value <- 15

  rand.row <- function() rnorm(num.steps, 1, step.std.dev)
  walk <- function () cumprod(rand.row()) * start.value
  all.walks <- t(replicate(100, walk()))
  plot(colMeans(all.walks), type = "l")

[I’m not an R guy, so that might not be the most typical style ever, but you get the idea...]

Aside: One of the things that I really don’t like about R coming from a background in other programming languages is that a function is evaluated before its arguments; i.e. an expression passed as a parameter to a function is not a value but is an implicit lambda, to be evaluated multiple times somewhere inside the function.

For example:

  replicate(2, rnorm(2))

creates a 2x2 matrix of independent random values. Whereas:

  x <- rnorm(2)
  replicate(2, x)

Instead creates a 2x2 matrix with a single random value for each row, repeated across all columns. And so if you want to decompose it, you need to do:

  x <- function() rnorm(2)
  replicate(2, x())

Which will re-call x twice inside the replicate function.

R does have call-by-need semantics, like (an impure) Haskell, but that's not why replicate's second argument gets evaluated multiple times:

  replicate <- function (n, expr, simplify = TRUE) 
     sapply(integer(n), 
        eval.parent(
           substitute(function(...) expr)), 
      simplify = simplify)

The replicate function uses R's compute-on-the-language to construct a new function that has expr as its body.

A closure's argument will be evaluated at most once.

Nice! I think you found the happy medium between the OPs two methods. His first is too long which creates more dependency on being able to read the language and his second is too short which creates more dependence on knowing the functions he's working with. It's very interesting to me as I study readability quite a bit: both long and short coding styles aren't as good as medium.

You're absolutely on track here

The way he did it is nice and minimal, and in 2 months when you go back to change something you will have no idea what you did.

I've had to do a few things in R and I always dread revisiting my code. Half the time it's easier and faster to just rewrite everything.

I've often felt that there is a world of terse, clear code waiting to be unlocked in R, but haven't been able to find it myself. This is a great example.

Had the same experience recently in Octave. Those moments of clarity are a big part of why I write software.

Slightly OT: I'd be really interested to hear a list of specific advantages to using R over using scipy. I have been holding off on learning R for years because scipy has served my purposes pretty well so far, but sometimes I wonder what I'm missing.

The extent of my experience with R was a couple of somewhat introductory statistics courses, so I’m not the best person to answer probably.

But I like dealing with numpy/scipy much, much more than R. Python as a language is I think much better designed, and numpy is a really nice tool for interacting with multidimensional arrays. When I write Python code, or read Python code written by anyone competent, I find program intent very easy to design/follow. Most of the R code I’ve seen “in the wild” is kind of a mess, because it is written by non-programmers many of whom have little experience or concept of code style. Additionally, as soon as a program has to do anything other than statistical analysis (examples: text munging, internet scraping, network communication, dealing with file formats, user interaction, etc.) Python is miles ahead.

The big advantages of R that I saw: (1) it has become the tool of choice in the academic statistics community, meaning that there is quite a lot of existing code for doing various sophisticated things, some of which you might have to implement yourself in Python, (2) it has some really nice graphing tools, (3) there seemed to be a few examples where a particular few lines of R code were more compact and clearer than the equivalent Python (can’t think of anything off-hand though).

I'm a pythonist - use it all the time. Except when I need to do some stats or graphing, and then it's R all the way. The R language is ugly, and warty, but it's awesomely powerful. Hell, the graphing library ggplot2 (http://had.co.nz/ggplot2/) is worth it by itself.

I'm not really a big fan of R but I have found it to be advantageous to use for some things. In particular, the big advantage of R is the huge user community. You can find a huge number of packages to use for many non-trivial tasks. Also the graphics tools are in my opinion superior to scipy.

Zen? This is madness!

In my mind, Matlab's (Octave's) array based programming makes sense. This? This does nothing that I expected it to do!

replicate seems to pretty randomly takes a function. Is that an R thing? I think what confuses me most is that there is nothing about this syntax that tells me that "cumprod (rnorm (1000, 1, 0.03))" hasn't already been evaluated! I could not, for the life of me, figure out why replicate didn't just create 100 exact copies. For example, why does replicate(...) evaluate, but the internals don't? This is driving me crazy!

This is slick...but:

How often do you want to generate random walks of this type where the variance of the process isn't dependent on its current level?

Observe that, as you increase the standard deviation of the random normal (to even .1), your "random walk" always walks to zero.

I don't mean to be a beady-eyed-pterodactyl but, as cool as clever one-liners sometimes are, often they solve toy problems. I say this as someone who loves R and uses it everyday and constantly is forced to brute-force with ugly inlined Rcpp code. (Which is fine)

That's an interesting point. In fact, the expectation of the random walk is always the same as its starting value, because, although most of the walks go to 0 like you observed, there are very occasionally walks that drift upward to astronomical values.

As for the usefulness of this kind of walk: the process we're modelling is an evolutionary one, where the rate of change is fixed (in this case within the species) and we'd like to detect 'random' (non-selected) evolutionary paths by comparing simulations to historical data.

In Matlab, where one is generally compelled (for better or worse) to think in terms of array operations, the obvious code would be:

plot(mean((cumprod(randn(100,1000) .* 0.03) .* 15))

Well, that is assuming one wants a line plot of the mean value of the "___location" at each time point across the population of walks. Personally, I find R's rather idiosyncratic approaches to data handling and function wrangling a bit hard to digest.

The most of this Zen comes from array paradigm, not functional; still it is nice to see another soul on the path to Renlightenment (-;

That's a matter of perspective. Folds and scans exist for all regular data types, not only arrays, so I consider this a part of functional programming and type theory.

Here he is computing the history of states of a random walk as scan (*) initial updates. If you wanted to simulate branching random walks, you'd work over the data type of multiway trees instead of arrays, but the algorithm to compute the history of intermediate states would otherwise be identical, using the scan function for multiway trees.

Those moments, its been a while since I've had one of those. (Must open compiler book)

How did that take a couple months? Even in R's slightly awkward language, random walks shouldn't be very difficult! I'm baffled, I must be missing something.

Writing the code did not take him several months; that's just the period of time he utilized this code. During that time, he tweaked it for evolving requirements. What took a couple of months was the path to his epiphany.

"modifying it on and off as research questions changed": he wasn't just writing that code.