Changes in version 7.0.2                        

Bug fixes

  - Exported and documented the GA selection helper functions required
    by the installed package namespace, including
    validate_selection_parameters().
  - Added regression coverage for the user-reported cd_mean,
    max_to_test, and mean_to_test workflows.

                        Changes in version 7.0.1                        

Bug fixes

  - subset_ga() / subset_ga_single() — "argument Test is missing" error
    fixed. Test now defaults to NULL in subset_ga(). The root cause was
    an R-language subtlety: assigning NULL to a named list element
    removes the key, so subset_ga_single() was silently dropping Test
    before forwarding the call via do.call(). Giving Test = NULL a
    default in subset_ga() makes both the no-test wrapper and any direct
    call without a test set work correctly.

  - Distance criteria (max_to_test, mean_to_test, neg_min_internal,
    neg_mean_internal) caused a "missing value where TRUE/FALSE needed"
    crash. These criteria were implemented in distances.R but were not
    registered in the unified criterion() dispatch function in
    criteria.R. Every call with a distance criterion was silently caught
    by tryCatch() and returned Inf for all fitness values, making
    improvement_rate = |Inf − Inf| / Inf = NaN, and NaN < tolconv
    evaluates to NA in R, triggering the error. The four distance
    criteria are now wired into the switch in criterion().

Notes for users upgrading from 5.x

  - The main entry-points are now subset_ga() (with test set) and
    subset_ga_single() (whole-population, no test set). The legacy
    functions GenAlgForSubsetSelection() and
    GenAlgForSubsetSelectionNoTest() are still available for backward
    compatibility.

  - Criterion names have changed from ALL-CAPS (CDMEAN2, PEVMEAN2, …) to
    snake_case (cd_mean, pev_mean, …). All old names are mapped via the
    unified criterion() function, so existing code continues to work.

  - Distance-based criteria (max_to_test, mean_to_test) require a square
    distance matrix as P (e.g., D in the demo data), not an MDS
    coordinate matrix (e.g., dmds). MDS / PCA coordinate matrices should
    be used with prediction-based criteria (cd_mean, pev_mean, etc.).

                        Changes in version 7.0.0                        

Major rewrite

  - New clean API: subset_ga(), subset_ga_single(), multi-objective
    variants.
  - Unified criterion() dispatcher covering classical (A, D, E, G, I
    optimality), PEV-based, CD-based, mixed-model, and distance-based
    criteria with full legacy name support.
  - Multi-criteria convergence detection with configurable window and
    restart mechanism.
  - Rank-based and hybrid selection, adaptive mutation, diversity
    preservation (crowding + fitness sharing).
  - Numerically stable matrix operations throughout
    (safe_matrix_inverse(), adaptive ridge).
  - Literature-corrected PEV and CD formulas based on Henderson (1984)
    and Fedorov (1972).

                 Changes in version 5.2.1 (2018-11-24)                  

  - Previous CRAN release.