Skip to contents

Train a MaxEnt Model with the Java-Equivalent Sequential Optimizer

Source: R/sequential.R
maxent_sequential_fit.Rd

Runs the full density.Sequential optimizer ported from the original Java Maxent 3.4.4 on a FeaturedSpace. Unlike maxent_fit(), which uses the historical goodAlpha-only loop, this trainer reproduces the real Java optimizer: feature selection via deltaLossBound, Newton step with 1-D line search (searchAlpha), every-10-iteration doParallelUpdate with undo on loss-violating batch steps, and per-feature state tracking.

Usage

maxent_sequential_fit(
  featured_space,
  max_iter = 500L,
  convergence = 1e-05,
  beta_multiplier = 1,
  min_deviation = 0.001,
  parallel_update_frequency = 10L,
  disable_convergence_test = FALSE,
  trajectory_iterations = integer(0)
)

Arguments

External pointer to a FeaturedSpace object (from maxent_featured_space()).

max_iter

Maximum number of iterations (default 500L).

convergence

Convergence threshold on the per-20-iteration loss improvement (default 1e-5). Ignored when disable_convergence_test = TRUE.

beta_multiplier

Regularization multiplier (default 1.0).

min_deviation

Minimum sample-deviation floor used in regularization (default 0.001).

parallel_update_frequency

Iteration frequency at which doParallelUpdate runs (default 10L, matching Java).

disable_convergence_test

Logical: when TRUE, the loop runs a fixed max_iter iterations with no early stop. This is required for deterministic per-iteration trajectory comparisons against the Java oracle. Default FALSE.

trajectory_iterations

Integer vector of 1-based iteration indices at which to capture (loss, entropy, lambdas) snapshots. May be empty. Snapshots outside [1, max_iter] are silently dropped.

Value

Named list with:

loss

Final regularized loss (scalar).

entropy

Shannon entropy of the trained distribution.

iterations

Number of training iterations completed.

converged

Logical: whether the convergence threshold was reached (always FALSE when disable_convergence_test = TRUE).

lambdas

Numeric vector of final lambda values.

trajectory

A data.frame with one row per captured checkpoint and columns iteration, loss, entropy, lambda_0, lambda_1, ..., lambda_{J-1}.

Details

The lambda trajectory produced by this trainer matches the Java oracle's to \(10^{-6}\) on \(\left\|\Delta\lambda\right\|_\infty\) at every checkpoint, on both the symmetric and asymmetric fixtures in maxentcppCompTest — see docs/ARCHITECTURE_xtensor_openmp.md and the Phase C baseline report for details.

See also

maxent_fit, maxent_featured_space.

Examples

# \donttest{
n   <- 100L
idx <- 90:99
env <- seq(0, 1, length.out = n)
f   <- maxent_linear_feature(env, "env1")
fs  <- maxent_featured_space(n, idx, list(f))
res <- maxent_sequential_fit(
    fs,
    max_iter                 = 500L,
    disable_convergence_test = TRUE,
    trajectory_iterations    = c(1L, 2L, 5L, 10L, 50L, 100L, 500L))
print(res$trajectory)
#>   iteration     loss  entropy   lambda_0
#> 1         1 4.411687 4.596430  0.4546303
#> 2         2 4.244694 4.572560  0.8845613
#> 3         5 3.868155 4.444896  2.0417866
#> 4        10 3.500539 4.185875  3.6185653
#> 5        50 2.783462 2.783462 16.6570831
#> 6       100 2.783462 2.783462 16.6570831
#> 7       500 2.783462 2.783462 16.6570831
# }