Caret leap forward for linear regression (feature selection) change nvmax - r

Question

I have been playing around with the leapForward method from the package leaps in conjunction with caret and found that it only provides 5 variables . according to the leaps package you can change nvmax to whatever number of subsets you wish.
I cannot seem where to fit this into the caret wrapper. I have tried putting it in the train statement, as well as creating an expand.grid line, and ti does not seem to work. Any help would be appreciated!
my code:
library(caret)
data <- read.csv(file="C:/mydata.csv", header=TRUE, sep=",")
fitControl <- trainControl(method = "loocv")
x <- data[, -19]
y <- data[, 19]
lmFit <- train(x=x, y=y,'leapForward', trControl = fitControl)
summary(lmFit)

Answer 1

The default behavior of caret is a random search over the tuning parameters.
You can specify a grid of parameters as you like, with the tuneGrid option.
Here is a reproducible example with the BloodBrain dataset. NB : I had to
transform the predictors with a PCA to avoid problems of multicolinearity
library(caret)
data(BloodBrain, package = "caret")
dim(bbbDescr)
#> [1] 208 134
X <- princomp(bbbDescr)$scores[,1:131]
Y <- logBBB
fitControl <- trainControl(method = "cv")
Default : random search of parameters
lmFit <- train(y = Y, x = X,'leapForward', trControl = fitControl)
lmFit
#> Linear Regression with Forward Selection
#>
#> 208 samples
#> 131 predictors
#>
#> No pre-processing
#> Resampling: Cross-Validated (10 fold)
#> Summary of sample sizes: 187, 188, 187, 187, 187, 187, ...
#> Resampling results across tuning parameters:
#>
#> nvmax RMSE Rsquared MAE
#> 2 0.6682545 0.2928583 0.5286758
#> 3 0.7008359 0.2652202 0.5527730
#> 4 0.6781190 0.3026475 0.5215527
#>
#> RMSE was used to select the optimal model using the smallest value.
#> The final value used for the model was nvmax = 2.
With a grid search of your choice.
NB : expand.grid is not necessary here. it is useful when you combine
several tuning parameters
lmFit <- train(y = Y, x = X,'leapForward', trControl = fitControl,
tuneGrid = expand.grid(nvmax = seq(1, 30, 2)))
lmFit
#> Linear Regression with Forward Selection
#>
#> 208 samples
#> 131 predictors
#>
#> No pre-processing
#> Resampling: Cross-Validated (10 fold)
#> Summary of sample sizes: 188, 188, 188, 186, 187, 187, ...
#> Resampling results across tuning parameters:
#>
#> nvmax RMSE Rsquared MAE
#> 1 0.7649633 0.07840817 0.5919515
#> 3 0.6952295 0.27147443 0.5250173
#> 5 0.6482456 0.35953363 0.4828406
#> 7 0.6509919 0.37800159 0.4865292
#> 9 0.6721529 0.35899937 0.5104467
#> 11 0.6541945 0.39316037 0.4979497
#> 13 0.6355383 0.42654189 0.4794705
#> 15 0.6493433 0.41823974 0.4911399
#> 17 0.6645519 0.37338055 0.5105887
#> 19 0.6575950 0.39628133 0.5084652
#> 21 0.6663806 0.39156852 0.5124487
#> 23 0.6744933 0.38746853 0.5143484
#> 25 0.6709936 0.39228681 0.5025907
#> 27 0.6919163 0.36565876 0.5209107
#> 29 0.7015347 0.35397968 0.5272448
#>
#> RMSE was used to select the optimal model using the smallest value.
#> The final value used for the model was nvmax = 13.
plot(lmFit)
Created on 2018-03-08 by the reprex package (v0.2.0).