autotune {autocart}

Find the best alpha, beta, and bandwidth values with k-fold cross-validation

Description

Find the best alpha, beta, and bandwidth values with k-fold cross-validation

Usage

autotune(
  response,
  data,
  locations,
  k = 8,
  control = NULL,
  customGroups = NULL,
  alphaVals = NULL,
  betaVals = NULL,
  bandwidthVals = NULL,
  outputProgress = FALSE,
  useSpatialNodes = FALSE,
  spatialNodesMethod = "idw",
  spatialNodesDistPower = 2,
  spatialNodesDistPowerRange = c(0, 2),
  spatialNodesModelByResidual = FALSE
)

Arguments

response

The vector of response values to test on.

data

The data frame of predictor variables.

locations

The n by 2 matrix of coordinate information for the known observations

k

The number of folds to create in k-fold cross-validation for tuning

control

An optional control function to send to the autocart creation function

customGroups

Here, you may supply custom groups for cross-validation. This parameter must be supplied as a factor and labeled from 1:numLevels.

alphaVals

Override the alpha values that are expanded in the grid in this function.

betaVals

Override the beta values that are expanded in the grid in this function.

bandwidthVals

Override the bandwidth values that are expanded in the grid in this function.

outputProgress

Print the result of the cross-validations as you are going. This is useful when the cross-validation will be very long and you do not wish to wait.

useSpatialNodes

Use an interpolative process at the terminal nodes of the autocart tree for the prediction process

spatialNodesMethod

The type of interpolation to use at the terminal nodes

spatialNodesDistPower

The power parameter to use in inverse distance weighting at terminal nodes

spatialNodesDistPowerRange

The ranged power parameter p1, p2 to use for a varying power parameter

spatialNodesModelByResidual

Do the interpolative process on the residuals of the prediction formed by response average at terminal nodes

Value

A list of the labeled optimal parameters that were chosen for the best predictive accuracy on cross-validation.

Examples

# Load some data for an autotune example
# (Note that a low sample size is used here for quick example computation.
#  In a practical application this function can be quite computationally
#  demanding due to the grid-search nature of the function.)
snow <- na.omit(read.csv(system.file("extdata", "ut2017_snow.csv", package = "autocart")))
y <- snow$yr50[1:35]
X <- data.frame(snow$ELEVATION, snow$MCMT, snow$PPTWT)[1:35, ]
locations <- as.matrix(cbind(snow$LONGITUDE, snow$LATITUDE))[1:35, ]

# Find optimal parameters via cross-validation. We'll search through the
# following alpha/beta/bandwidth values:
alphaVec <- c(0.0, 0.5)
betaVec <- c(0.0, 0.2)
bandwidthVec <- c(1.0)

# We'll find the optimal values with 3-fold cross validation:
# (Due to the large number of cross-validations and trainings that occur,
# this can take a few minutes.)
myTune <- autotune(y, X, locations, k = 3, alphaVals = alphaVec,
                   betaVals = betaVec, bandwidthVals = bandwidthVec)
# Inspect the results
myTune