gbt.train {agtboost}  R Documentation 
gbt.train
is an interface for training an agtboost model.
gbt.train(
y,
x,
learning_rate = 0.01,
loss_function = "mse",
nrounds = 50000,
verbose = 0,
gsub_compare,
algorithm = "global_subset",
previous_pred = NULL,
weights = NULL,
force_continued_learning = FALSE,
offset = NULL,
...
)
y 
response vector for training. Must correspond to the design matrix 
x 
design matrix for training. Must be of type 
learning_rate 
control the learning rate: scale the contribution of each tree by a factor of 
loss_function 
specify the learning objective (loss function). Only prespecified loss functions are currently supported.

nrounds 
a justincase max number of boosting iterations. Default: 50000 
verbose 
Enable boosting tracing information at ith iteration? Default: 
gsub_compare 
Deprecated. Boolean: Globalsubset comparisons. 
algorithm 
specify the algorithm used for gradient tree boosting.

previous_pred 
prediction vector for training. Boosted training given predictions from another model. 
weights 
weights vector for scaling contributions of individual observations. Default 
force_continued_learning 
Boolean: 
offset 
add offset to the model g(mu) = offset + F(x). 
... 
additional parameters passed.

These are the training functions for an agtboost.
Explain the philosophy and the algorithm and a little math
gbt.train
learn trees with adaptive complexity given by an information criterion,
until the same (but scaled) information criterion tells the algorithm to stop. The data used
for training at each boosting iteration stems from a second order Taylor expansion to the loss
function, evaluated at predictions given by ensemble at the previous boosting iteration.
An object of class ENSEMBLE
with some or all of the following elements:
handle
a handle (pointer) to the agtboost model in memory.
initialPred
a field containing the initial prediction of the ensemble.
set_param
function for changing the parameters of the ensemble.
train
function for retraining (or from scratch) the ensemble directly on vector y
and design matrix x
.
predict
function for predicting observations given a design matrix
predict2
function as above, but takes a parameter max number of boosting ensemble iterations.
estimate_generalization_loss
function for calculating the (approximate) optimism of the ensemble.
get_num_trees
function returning the number of trees in the ensemble.
Berent Ånund Strømnes Lunde, Tore Selland Kleppe and Hans Julius Skaug, "An Information Criterion for Automatic Gradient Tree Boosting", 2020, https://arxiv.org/abs/2008.05926
## A simple gtb.train example with linear regression:
x < runif(500, 0, 4)
y < rnorm(500, x, 1)
x.test < runif(500, 0, 4)
y.test < rnorm(500, x.test, 1)
mod < gbt.train(y, as.matrix(x))
y.pred < predict( mod, as.matrix( x.test ) )
plot(x.test, y.test)
points(x.test, y.pred, col="red")