Statistics 4240 - Data Mining

Here's the CV code we looked at tonight. I believe there was
something else I said I would post to the blog but now I can't remember...

  numFolds <- 3      # how many of the possible 10 folds that are actually done
  candidateBounds <- c(0.0001,0.001,0.005,0.01,0.03,0.07,0.1)
  numCandidateBounds <- length(candidateBounds)
  cvRMSE <- rep(0,numCandidateBounds)
 
  cvIndex <- rep(1:10,trunc(nTrain/10)+1)[1:nTrain]   # 1-10 repeated the rigtt # of times
  cvIndex <- sample(cvIndex)                          # then scrambled
  for (j in 1:numFolds) {
    for (k in (1:numCandidateBounds)) {
 #     cat("fold:",j," candidate:",k,"\n")
      mylm<-l1ce(y ~ .,data=fooTrain[cvIndex!=j,],bound=candidateBounds[k])
      myPred<-predict(mylm,fooTrain[cvIndex==j,])
      cvRMSE[k] <- cvRMSE[k] + sqrt(mean((fooTrain[cvIndex==j,]$y - myPred)^2))
    }
  }
  k=1;
  while ((cvRMSE[k] > (min(cvRMSE)-0.0000001)) && (k < numCandidateBounds)) {
    k <- k+1;
  }
  if (k > 1) {k <- k-1}   # back off one from the best
  cat(candidateBounds[k],"\n")