glue_ann: Train a GLUE-ANN model ensemble

In rogiersbart/RGLUEANN: Non-linear probabilistic data-driven modelling using general likelihood uncertainty estimation and artificial neural networks

Description

glue_ann trains a GLUE-ANN model ensemble. A dataframe has to be provided with the independent variable(s), and a vector with the dependent variable. A list of parameters has to be set to define the a priori distributions of the stochastic parameters, and fixed values for the ones considered to be constant.

Usage

glue_ann(inputData, outputData, orthogonal = "Cor",
  variableNumber = c(1:ncol(inputData)), dimRed = T, rescale.output = T,
  ES = T, randomESPart = 0.5, nonEqualMeans = T,
  maxMeanDiff = mean(outputData) * 0.2, nCycles = 100, HLrange = c(1:1),
  HNrange = c(1:10), maxHNPerVar = Inf, hidden.layer = "tansig",
  output.layer = "sigmoid", nSets = 100, nStedinger = 50,
  weighting = "Stedinger", cv = F)

Arguments

inputData	input data frame
outputData	output data frame
orthogonal	character; orthogonalize the input data; options are: 'Cov', 'Cor', and 'No'; default: 'Cor'
variableNumber	number of input variables; provide vector for stochastic treatment; defaults to number of columns of inputData
dimRed	logical; TRUE selects always most variance explaining PCs; FALSE chooses PCs randomly; default: TRUE
rescale.output	logical; rescale output to [0,1] range; default: TRUE
ES	logical; enable early stopping; default: TRUE
randomESPart	part of the data used for early stopping; default: 0.5
nonEqualMeans	force similar mean for the early stopping samples as for the training samples; default: TRUE
maxMeanDiff	maximum difference between early stopping and training observation means; defaults to 20 percent of the target data mean
nCycles	maximum number of training cycles, provide a vector for stochastic treatment; default: 100
HLrange	amount of hidden layers, provide vector for stochastic treatment; default: c(1:1)
HNrange	amount of hidden nodes, provide vector for stochastic treatment; default: c(1:10)
maxHNPerVar	maximum amount of hidden nodes per input variable; default: Inf
hidden.layer	hidden layer activation function; options are 'tansig', 'sigmoid', 'purelin'; default: 'tansig'
output.layer	output layer activation function; options are 'tansig', 'sigmoid', 'purelin'; default: 'sigmoid'
nSets	number of models in the ensemble; default: 100
nStedinger	number of random variates for weights based on Stedinger et al. (2008); default: 50
weighting	GLUE weighting procedure; options are 'MSE', 'MSEes' (for weights based on the early stopping subset performance), 'NSEff' and 'Stedinger'; default: 'Stedinger'
cv	logical; do leave-one-out cross-validation; default: FALSE

Value

Object of class glue_ann

rogiersbart/RGLUEANN documentation built on May 27, 2019, 12:16 p.m.