CODEWITHSUNDEEP
rrandom-forestr-caret
Rene Recktenwald
Rene Recktenwald

Reputation: 175

How can I add more trees to a random forest in R

Is there a canonical way to iteratively add trees to a random forest? Let's say I am using the caret package and I use something like

rf_fit <- train(y~.,data=df,method="rf",ntree = N)

for some N and then I would like to continue adding trees to it, how would I go about that?

Upvotes: 0

Views: 208

Answers (1)

Elia
Elia

Reputation: 2584

You could create your own function and lapply across ntree:

data <- iris

fit_tree <- function(ntree){
rf_fit <- train(Species~.,data=iris,method="rf",ntree = ntree)
}

fit <- lapply(seq(100,500,100),fit_tree)

Here fit is a list of 5 randomForests model each fitted with the ntrees specified in the first argument of lapply. I don't know if is possible to add trees to the same model. If the model fitted with n trees don't reach the accuracy you want, you can simply re-fit the model with n+100 trees for example (but keep in mind that increasing the number of trees doesn't necessarily improve the accuracy, indeed, it could worsen performance. In the caret package the default ntrees is 500 as suggested by Breiman in his original paper (Breiman, 2001)).

EDIT

to add trees to an existing randomForests model you can do something like this:

fit_tree <- function(how.many){
rf_fit <- randomForest(Species~.,data=iris)
new_fit <- grow(rf_fit,how.many)
}

p <-lapply(seq(10,100,10),fit_tree)

Here the starting ntree is the default one (I. e. 500) and the lapply adds 10 trees for each iteration. With this approach is not so helpful to tune the mtry parameter with caret since the best parameter values would be found only for the first call to randomForest, but not for the updated model

Upvotes: 1

Related Questions