The Art of R Programming

0.012 0.000 0.015

> system.time(

+{

+ c <- 0

+ for (i in 1:length(x))

+ if (x[i] %% 2 == 1) c <- c+1

+ return(c)

+}

+)

user system elapsed

0.308 0.000 0.310

You might wonder whether it matters in this case, since even the loop

version of the code took less than a second to run. But if this code had been

part of an enclosing loop, with many iterations, the difference could be

important indeed.

Examples of other vectorized functions that may speed up your code

areifelse(),which(),where(),any(),all(),cumsum(), andcumprod(). In the

matrix case, you can userowSums(),colSums(), and so on. In “all possible

combinations” types of settings,combin(),outer(),lower.tri(),upper.tri(),

orexpand.grid()may be just what you need.

Thoughapply()eliminates an explicit loop, it is actually implemented in

R rather than C and thus will usually not speed up your code. However, the

otherapplyfunctions, such aslapply(), can be very helpful in speeding up

your code.

14.2.2 Extended Example: Achieving Better Speed in a Monte Carlo

Simulation

In some applications, simulation code can run for hours, days, or even

months, so speedup methods are of high interest. Here, we’ll look at two

simulation examples.

To begin, let’s consider the following code from Section 8.6:

sum <- 0

nreps <- 100000

for (i in 1:nreps) {

xy <- rnorm(2) # generate 2 N(0,1)s

sum <- sum + max(xy)

}

print(sum/nreps)

Here’s a revision (hopefully faster):

nreps <- 100000

xymat <- matrix(rnorm(2*nreps),ncol=2)

308 Chapter 14