Behavior with small numbers of trees

[djhwolters]

I am using ranger to predict threshold exceedance probabilities using QRF. I am puzzled by the behaviour with small numbers of trees. With for example 100 trees, I get nice continuous probabilities. But for 5 trees, it predicts either 0, 0.2, 0.4, 0.6, 0.8 or 1.0, and for 3 trees 0, 0.5 or 1. It looks like the individual trees predict deterministically, and the probabilities are derived by combining these deterministic trees. But for QRF as I understand, the probability distribution should come from the ecdf of all data points in the terminal nodes, so even for 1 tree, many different probabilites should be possible. Any ideas on this?

Code that shows this behaviour:

`library(ranger)

#load data set
data(mtcars)

#extract 1 column as test forecast row
xtest=mtcars[3,]
mtcars=mtcars[-c(3)]

#fit model using 50 trees
model <- ranger(data=mtcars,quantreg=TRUE,dependent.variable.name='mpg',importance='permutation',classification=FALSE,num.trees=50,min.node.size=2)
outcomes <- predict(model,xtest,what=function(x) ecdf(x)(23),type='quantiles')
print(outcomes$predictions)

#fit model using 3 trees
model <- ranger(data=mtcars,quantreg=TRUE,dependent.variable.name='mpg',importance='permutation',classification=FALSE,num.trees=3,min.node.size=2)
outcomes <- predict(model,xtest,what=function(x) ecdf(x)(23),type='quantiles')
print(outcomes$predictions)

#fit model using 1 tree
model <- ranger(data=mtcars,quantreg=TRUE,dependent.variable.name='mpg',importance='permutation',classification=FALSE,num.trees=1,min.node.size=2)
outcomes <- predict(model,xtest,what=function(x) ecdf(x)(23),type='quantiles')
print(outcomes$predictions)`

[mnwright]

The reason for this is the speedup discussed here: lorismichel/quantregForest#3.

Maybe we should add a note to the documentation?