Timing Test Results On Earth Version 0.1
----------------------------------------

Tested on a 1.5GHz Dell laptop with the robot arm from Friedman's Fast MARS paper.

N     ncol(x) fast.k  degree     time    gcv             grsq  nterms  npreds
3000     5       -1      10      4.9     0.0025          0.98    54      4       original
3000     5       20      10      3.6     0.0024          0.98    59      4       original
3000     30      -1      10      41      0.0025          0.98    54      4       original
3000     30      20      10      27      0.0024          0.98    59      4       original

3000     5       -1      10      5       0.0025          0.98    54      4       no -O3 (i.e. -O2)
3000     5       20      10      3.7     0.0024          0.98    59      4       no -O3 (i.e. -O2)
3000     30      -1      10      41      0.0025          0.98    54      4       no -O3 (i.e. -O2)
3000     30      20      10      27      0.0024          0.98    59      4       no -O3 (i.e. -O2)

3000     5       -1      10      4.9     0.0025          0.98    54      4       no INLINE
3000     5       20      10      3.7     0.0024          0.98    59      4       no INLINE
3000     30      -1      10      41      0.0025          0.98    54      4       no INLINE
3000     30      20      10      27      0.0024          0.98    59      4       no INLINE

3000     5       -1      10      6.7     0.0025          0.98    54      4       no BLAS
3000     5       20      10      4.8     0.0024          0.98    59      4       no BLAS
3000     30      -1      10      62      0.0025          0.98    54      4       no BLAS
3000     30      20      10      39      0.0024          0.98    59      4       no BLAS

3000     5       -1      10      5.7     0.0025          0.98    54      4       no BETA_CACHE
3000     5       20      10      4.2     0.0024          0.98    59      4       no BETA_CACHE
3000     30      -1      10      49      0.0025          0.98    54      4       no BETA_CACHE
3000     30      20      10      32      0.0024          0.98    59      4       no BETA_CACHE

3000     5       -1      10      7.6     0.0025          0.98    54      4       no INLINE, no USE_BLAS, no BETA_CACHE
3000     5       20      10      5.4     0.0024          0.98    59      4       no INLINE, no USE_BLAS, no BETA_CACHE
3000     30      -1      10      70      0.0025          0.98    54      4       no INLINE, no USE_BLAS, no BETA_CACHE
3000     30      20      10      44      0.0024          0.98    59      4       no INLINE, no USE_BLAS, no BETA_CACHE

Notes
-----

1. INLINE appears to make no difference for this model -- is gcc actually doing the inlining?
   Using the Microsoft compiler, INLINE makes a difference(not shown here)

2. -O3 appears to make no difference for this model

3. fast.k gives less improvement than expected from the Fast MARS paper

4. To see a malloc fail (in RegressAndFix) on a 512 MB machine: test(s, 100000, -1,  1, 5)
   This probably means that the peak memory use occurs in RegressAndFix,
   which could be fixed quite easily but I haven't done so yet.

Here is the code used to produce the above table.

robotArm <- function(x) {
  x. <- with(x, l1 * cos(theta1) - l2 * cos(theta1 + theta2) * cos(phi))
  y  <- with(x, l1 * sin(theta1) - l2 * sin(theta1 + theta2) * cos(phi))
  z  <- with(x, l2 * sin(theta2) * sin(phi))
  sqrt(x.^2 + y^2 + z^2)
}
test <- function(s, N, fast.k, degree, ndummy) {
  set.seed(1)   # for reproducibility
  gc()
  l1     <- runif(N, 0, 1)
  l2     <- runif(N, 0, 1)
  theta1 <- runif(N, 0, 2 * pi)
  theta2 <- runif(N, 0, 2 * pi)
  phi    <- runif(N, -pi/2, pi/2)
  x <- cbind(l1, l2, theta1, theta2, phi)
  if (ndummy > 0) for (i in 1:ndummy)
     x <- cbind(x, runif(N, 0, 1))
  x <- data.frame(x)
  e.time <- system.time(e <- earth(x, robotArm(x), degree=degree, nk=201, fast.k=fast.k))
  options(digits=2)
  cat(N, "\t", ncol(x), "\t", fast.k, "\t", degree, "\t", e.time[1], "\t",  e$gcv,
      "\t",  e$grsq, "\t",  length(e$selected.terms), "\t",
     get.nused.preds.per.subset(e$dirs, e$selected.terms), "\t", s, "\n")
}
s <- "original"
cat("N     ncol(x) fast.k  degree     time    gcv             grsq   nterms  npreds\n")
test(s, 3000, -1,  10, 0)
test(s, 3000, 20,  10, 0)
test(s, 3000, -1,  10, 25)
test(s, 3000, 20,  10, 25)

Forden Wales Mar 2007
