Begin A Greedy Multiple Kernel Design Search

This method creates an object of type greedy_multiple_kernel_experimental_design and will immediately initiate a search through allocation space for forced balance designs. For debugging, you can use set the seed parameter and num_cores = 1 to be assured of deterministic output.

Usage

initGreedyMultipleKernelExperimentalDesignObject(
  X = NULL,
  kernel_names = NULL,
  Kgrams = NULL,
  kernel_weights = NULL,
  objective = "added_pct_reduction",
  maximum_gain_scaling = TRUE,
  nT = NULL,
  max_designs = 10000,
  kernel_pre_num_designs = 1000,
  wait = FALSE,
  start = TRUE,
  max_iters = Inf,
  semigreedy = FALSE,
  diagnostics = FALSE,
  num_cores = 1,
  seed = NULL,
  verbose = TRUE,
  use_safe_inverse = FALSE
)

Arguments

X

The design matrix with $n$ rows (one for each subject) and $p$ columns (one for each measurement on the subject). This parameter must be specified unless you choose objective type "kernel" in which case, the Kgrams parameter must be specified.

kernel_names

A vector of M >= 1 strings indicating the kernels to be used. Valid values are "mahalanobis", "gaussian", "laplacian", "exponential", "inv_mult_quad", "poly_2", "poly_3". Or a special value "manual" to indicate the kernels are specified manually using the Kgrams parameter.

Kgrams

A list of M >= 1 elements where each is a n x n matrix whose entries are the evaluation of the kernel function between subject i and subject j. Default is NULL.

kernel_weights

A vector of M weights for each kernel which should sum to 1. Default is NULL which specifies equal weighting.

objective

The method used to aggregate the kernel objective functions together. Default is "added_pct_reduction".

maximum_gain_scaling

Should we scale the kernels to have the same maximum gain? Default is TRUE.

nT

The number of treatments to assign. Default is NULL which is for forced balance allocation i.e. nT = nC = n / 2 where n is the number of rows in X (or Kgram if X is unspecified).

max_designs

The maximum number of designs to be returned. Default is 10,000. Make this large so you can search however long you wish as the search can be stopped at any time by using the stopSearch method.

kernel_pre_num_designs

The number of initial designs to search through before starting the greedy search for each kernel. Default is 1000.

wait

Should the R terminal hang until all max_designs vectors are found? The deafult is FALSE.

start

Should we start searching immediately (default is TRUE).

max_iters

The maximum number of iterations of the greedy search algorithm to run. Default is Inf.

semigreedy

Should we use a fully greedy approach or the quicker semi-greedy approach? The default is FALSE corresponding to the fully greedy approach.

diagnostics

Should the objective function values at each iteration be saved? Default is FALSE.

num_cores

The number of CPU cores to use for the search. Default is 1.

seed

The set to set for deterministic output. This should only be set if num_cores = 1 otherwise the output will not be deterministic. Default is NULL for no seed set.

verbose

Should the algorithm emit progress output? Default is TRUE.

use_safe_inverse

Should a regularized inverse be used for the Mahalanobis objective? Default is FALSE.

Value

An object of type greedy_multiple_kernel_experimental_design which can be further operated upon

Author

Adam Kapelner

Examples

 if (FALSE) { # \dontrun{
  library(MASS)
  data(Boston)
 #pretend the Boston data was an experiment setting 
  #first pull out the covariates
 X = Boston[, 1 : 13]
 #begin the greedy design search
  mk = initGreedyMultipleKernelExperimentalDesignObject(X, 
    max_designs = 100, num_cores = 3, kernel_names = c("mahalanobis", "gaussian"))
  #wait
  res = resultsMultipleKernelGreedySearch(mk, max_vectors = 2)
  design = res$ending_indicTs[1, ] #ordered already by best-->worst
 design
  #how far have we come of the 100 we set out to do?
  mk
  #we can cut it here
  stopSearch(mk)
  } # }