Automatically tunes Random Forest

Question

I want to automatically tune the Random Forest Model as my variables keeps on changing on real time basis. I am using the iris dataset.

My code is

jk = iris

library(randomForest)
library(caret)

fgl.res <- tuneRF(lm[,-5], lm[,5], stepFactor=1.5)

o/p:
mtry = 2  OOB error = 5.33% 
Searching left ...
Searching right ...
mtry = 3    OOB error = 4% 
0.25 0.05 
mtry = 4    OOB error = 5.33% 
-0.3333333 0.05 

Above I want that in fgl.res code it automatically chooses the Species column index ie 5 in iris dataset and insert in fgl.res code.

Then to use a first line of fgl.res output(o/p) and takes value of "mtry = 2 OOB error = 5.33% " and use it in random forest code, ie assign value to mtry and oob.error as shown below:

mod2<-randomForest(Species~., data=lm, ntree=50, mtry=2, oob.error=0.0533)

I have tried many ways but nothing work out on how to automatically insert value into code from fgl.res output.

Answer 1

I dont know if I correctly understood your problem, but you might use this approach.
When you use tuneRF you have to choose mtry with the lowest OOB error. I use invisible(capture.output(...)) function to hide display any output in console after entering tuneRF function.

Example:

# load library
library(randomForest)
library(caret)

# data
data_iris = iris

# repeat the analysis
set.seed(4543)

# tuneRF 
invisible(capture.output(fgl.res <- tuneRF(x = data_iris[,-5], y= 
data_iris[,5], stepFactor=1.5)))

# choose the best mtry based on the lowest OOB error
best_mtry <- fgl.res[fgl.res[, 2] == min(fgl.res[, 2]), 1]

# choose the lowest OOB error
best_oob  <- fgl.res[fgl.res[, 2] == min(fgl.res[, 2]), 2]

# caluclate RF
mod2<-randomForest(Species~., data=data_iris, ntree=50, mtry=best_mtry,
oob.error=best_oob)

When you want to just extract a first line of fgl.res output and take value of mtry and OOB error, you have to use:

# choose the best mtry based on the lowest OOB error
best_mtry <- fgl.res[1, 1]

# choose the lowest OOB error
best_oob  <- fgl.res[1, 2]

I know that you posted this question 10 months ago, but maybe this approach would be useful for other users.