Optimal training set determination

This function is designed for determining optimal training set.

Usage

optTrain(
  geno,
  cand,
  n.train,
  subpop = NULL,
  test = NULL,
  method = "rScore",
  min.iter = NULL,
  console = TRUE,
  seed = NULL,
  lambda = NULL,
  threads = NULL
)

Arguments

geno: A numeric matrix of principal components (rows: individuals; columns: PCs).
cand: An integer vector of which rows of individuals are candidates of the training set in the geno matrix.
n.train: The size of the target training set. This could be determined with the help of the ssdfgp function provided in this package.
subpop: A character vector of sub-population's group name. The algorithm will ignore the population structure if it remains NULL. Without test, it has one entry per candidate (same order as cand); with test, it has one entry per row of geno.
test: An integer vector of which rows of individuals are in the test set in the geno matrix. The algorithm will use an un-target method if it remains NULL.
method: A criterion name; see criteria for the list. "rScore" (default), "PEV" and "CD" are the criteria of v2.4.2.
min.iter: Minimum iteration of all methods can be appointed. One should always check if the algorithm is converged or not. A minimum iteration will set by considering the candidate and test set size if it remains NULL.
console: Default: TRUE. Set it to FALSE if you don't want the function printing out the number count of each iteration.
seed: Default: NULL. An integer passed to set.seed before the search, which changes the global random-number state. Leave it NULL to use the current state.
lambda: Default: NULL. A positive ridge penalty; NULL uses the default of each criterion (see criteria).
threads: Default: NULL (one thread). The number of threads that Eigen may use inside the matrix products of one search (OpenMP). For the exchange search it only helps when geno has many columns (several hundred; about 1.3 to 1.5 times faster at 404 columns on 2 threads) and the training set found is the same apart from rounding in the last digits. For the targeted r-score and PEV search without sub-populations (a genetic algorithm), the 30 candidate sets of each generation are scored in parallel, and the result does not depend on threads. To run many independent searches in parallel use SSDFGS.

Value

This function will return a list with OPTtrain (a vector of chosen optimal training set), TOPscore (highest scores of before iteration), ITERscore (criteria scores of each iteration), n.iter (number of iterations run), and converged (TRUE if the search stopped by the convergence rule, FALSE if it stopped at twice min.iter).

Details

All twelve search routes of v2.4.2 (with or without sub-populations, targeted or un-targeted, and the three criteria) are kept. The search moves, their random-number order, the default numbers of iterations, and the stopping rule are the same as in v2.4.2, so that a given seed returns the training set of v2.4.2 (apart from the float-to-double change of the scores, see Details of r_score). The r-score is the criterion of Ou and Liao (2019) <doi:10.1007/s00122-019-03387-0>.

Only the r-score, PEV, and CD are tested carefully: they reproduce the output recorded from v2.4.2 on all twelve routes. The 15 criteria added in v3 follow their papers and run on all routes, but they are not carefully tested; evaluate them on your own data before relying on them.

References

Ou, J.-H., & Liao, C.-T. (2019). Training set determination for genomic selection. Theoretical and Applied Genetics, 132, 2781-2792. <doi:10.1007/s00122-019-03387-0>

Author

Jen-Hsiang Ou

Examples

data(geno)
if (FALSE) optTrain(geno, cand = 1:404, n.train = 100) # \dontrun{}