bamlss (version 1.0-0)

n: Neural Networks for BAMLSS

Description

This smooth constructor implements single hidden layer neural networks with sigmoidal activation functions.

Usage

## The neural network smooth constructor.
n(..., k = 10)
## Initialize weights.
n.weights(nodes, k, r = NULL, s = NULL, type = c("sigmoid", "gauss"), ...)
## For mgcv.
# S3 method for nn.smooth.spec
smooth.construct(object, data, knots, ...)

Arguments

…

For function n() a formula of the type ~x1+x2+x3 that specifies the covariates that should be modeled by the neural network. For function boost.nnet.predict() arguments passed to function predict.bamlss.

k

For function n(), the number of hidden nodes of the network. Note that one can set an argument split = TRUE to split up the neural network into, e.g., nsplit = 5 parts with k nodes each. For function nnet.weights(), argument k is the number of input variables of the network (number of covariates).

nodes

Number of nodes for each layer, i.e., can also be a vector.

r, s

Parameters controlling the shape of the activation functions.

type

The type of activation function that should be used.

object, data, knots

See smooth.construct.

Value

Function n(), similar to function s a simple smooth specification object.

See Also

bamlss, predict.bamlss, bfit, boost

Examples

# NOT RUN {
## Simulate data.
set.seed(123)
d <- GAMart()

## Estimate model.
f <- num ~ n(x1,k=50) + n(x2,k=50) + n(x3,k=50) + n(~lon+lat,k=50,split=TRUE)

## Set the seed, weights are initialized randomly.
set.seed(111)
b <- bamlss(f, data = d, sampler = FALSE)

## Plot estimated effects.
plot(b)

## Boosted version.
set.seed(111)
b <- bamlss(f, data = d, sampler = FALSE, optimizer = boost,
  nu = 0.05, maxit = 4000)

plot(b)

## Second example with MCMC.
set.seed(222)
n <- 2000
d <- data.frame("x1" = runif(n, 0, pi), "x2" = runif(n, 0, pi))
d$y <- sin(d$x1) * sin(d$x2) + rnorm(n, sd = exp(-2 + sin(2 * d$x2)))

f <- list(
  y ~ n(~x1+x2,k=50),
  sigma ~ n(x2)
)

b <- bamlss(f, data = d)
plot(b, ask = FALSE)

## Initialize some weights.
set.seed(111)
n.weights(10, 3)
# }