Optimize csmoothEM initialization by comparing multiple methods after a warm start

Runs several initialization methods (via parallel_initial_csmoothEM), each for num_iter warm-start iterations, then selects the best fit.

Selection criterion:

  • If adaptive == "ml", selects the fit with the largest ml_last (collapsed objective \(\mathcal{C}\)).

  • Otherwise, selects the fit with the largest elbo_last.

If plot=TRUE, plots traces across methods and highlights the selected best method. When adaptive=="ml", the primary plot shows ml_trace; otherwise it shows elbo_trace. If two_panel=TRUE, a second panel shows loglik_trace.

optimize_initial_csmoothEM(
  X,
  methods = c("PCA", "tSNE", "random", "fiedler", "pcurve"),
  num_iter = 5,
  num_cores = 2,
  K = NULL,
  adaptive = "prior",
  lambda_min = 1e-08,
  lambda_max = 1e+08,
  sigma_update = c("mstep", "ml"),
  sigma_min = 1e-10,
  sigma_max = 1e+10,
  plot = FALSE,
  two_panel = FALSE,
  seed = NULL,
  quiet = TRUE,
  ...
)

Arguments

X

Numeric matrix (n x d).

methods

Character vector of initialization methods. Default includes "PCA", "tSNE", "random", "fiedler", "pcurve", "isomap".

num_iter

Integer \(\ge 1\). Number of warm-start iterations to run per method.

num_cores

Integer \(\ge 1\). Number of cores to use (see parallel_initial_csmoothEM).

K

Optional integer \(\ge 2\). Number of mixture components. If NULL, the default logic in initialize_csmoothEM is used.

adaptive

Character. One of "none", "prior", "ml". Logical values are accepted for backward compatibility: TRUE is treated as "prior" and FALSE as "none".

lambda_min, lambda_max

Positive bounds for lambda_vec when adaptive!="none".

sigma_update

Character. Only used when adaptive="ml" (passed through). Either "mstep" or "ml".

sigma_min, sigma_max

Positive bounds for sigma2 when adaptive="ml" and sigma_update="ml".

plot

Logical; if TRUE, plot traces across methods.

two_panel

Logical; if TRUE, also plot penalized objective traces in a second panel.

seed

Optional integer seed. If provided, a different derived seed is used per method.

quiet

Logical; reserved for future use.

...

Additional arguments passed to initialize_csmoothEM (and downstream ordering routines).

Value

A csmooth_em object corresponding to the selected best method. The returned object has an additional meta$initial_search list containing:

  • best_method: selected method name

  • criterion: selection criterion ("ml_last" or "elbo_last")

  • summary: the per-method summary data.frame

  • fits: the full named list of fits

  • options: options used for the search