som.tunecontrol: Control parameters for the som.tune function

In yasomi: Yet Another Self Organising Map Implementation

Description

Creates a list of parameters for the som.tune function.

Usage

som.tunecontrol(somgrid, init = "pca", ninit = 1, assignment = "single",
                radii = c(2, 2/3 * somgrid$diam), nradii = 10,
                innernradii = 30, maxiter = 75, annealing = "power",
                kernel = "gaussian", criterion = error.quantisation)

Arguments

somgrid	an object of class "somgrid"
init	prototypes initialization method. Valid values are "pca" and "random". The former corresponds to principal component based initialization (see sominit.pca), while the latter uses randomly selected observations as initial values for the prototypes (see sominit.random)
ninit	number of initial prototype values to test (only relevant for init="random")
assignment	assignment method with valid values "single" and "heskes" (see batchsom)
radii	the range of radii to explore, i.e., a vector of length two containing a minimal and a maximal value of radii. The default minimum radius is 2 (almost purely local k-means like optimization) while the maximum is equal to two third of the diameter of the prior struture
nradii	number of radii to generate from the range specified in radii
innernradii	number of radii to use in the annealing scheme during the SOM fitting (see batchsom)
maxiter	maximal number of iteration for each radius during fitting (see batchsom)
annealing	annealing scheme with valid values "power" (exponential like annealing) and "linear" (linear scheme)
kernel	kernel chosen between "gaussian" and "linear"
criterion	an error criterion, i.e., a function that evaluate the quality of a fitted som on a dataset

Details

The parameters init, assignment, annealing and kernel can contain a list of values rather than a single value. In this case, all combinations are compared in som.tune. For instance, if all parameters have default values expect for kernel=c("gaussian","linear") then som.tune will fit 10 SOM with a gaussian kernel as well as 10 SOM with a linear kernel and select the best one.

The error criterion is a function which takes a object returned by the appropriate batchsom method (e.g. a "somnum" object for standard vector data and a "relationalsom" object for dissimilarity data) and returns a numerical value. som.tune optimises for small values of this criterion.

Value

a list containing all the above parameters with one exception: 'radii' contains a vector of radii to test rather than the bounds specified in the function call.

Note

this function is modelled after David Mayer's tune.control function.

Author(s)

Fabrice Rossi

See Also

som.tune

yasomi documentation built on May 2, 2019, 5:59 p.m.