Language: en-US), inst/WORDLIST for
technical terms and names; spelling::spell_check_package()
reports no errors. No change to code behavior.MultKAN.fit(opt = "Adam"), 200-300 steps, fixture
pykan_trajectory.json, test
test-pykan-training.R). Without grid updates the loss
trajectory and final parameters agree to 1e-14. With pykan’s grid-update
schedule the remaining difference came only from the ridge penalty of
the coefficient refit; new argument kan_fit(grid_ridge = )
(default 1e-8, unchanged behavior), and
grid_ridge = 0 (plain least squares, as pykan) agrees to
1e-13. kan_update_grid(ridge = 0) now uses a QR
least-squares solve.inst/validation/compare_pykan_training.R.inst/benchmarks/): B-spline KAN
against OLS, GAM, random forest, ANN and a Fourier KAN comparator, same
data and out-of-sample protocol, 10 seeds.inst/applications/): base R PLS-PM
(checked against plspm), fold-internal latent scores and repeated
cross-validation of KAN vs. the PLS-SEM structural model on three public
customer data sets (ECSI mobi,
plspm::satisfaction, corp_rep_data).
pls_pm() supports formative (mode B) blocks.inst/REPRODUCE.md lists every script, seed and software
version behind the reported results.plot.kan_pd() failed when xlab or
ylab was supplied.MultKAN (architectures 2-3-1, 3-4-2-1, 2-5-3-2): forward
pass, loss and all layer gradients agree to < 1e-10, five full-batch
Adam steps follow the same trajectory (< 1e-10), and the network-wide
sample-based grid update places the same knots (layer 1 exactly, deeper
layers < 4e-6, the difference coming from least squares vs. ridge
refits). Generator script:
inst/validation/pykan_multilayer_reference.py.cran-comments.md, MIT license files,
cph role, Language field.R CMD check --no-manual 1 NOTE (fixed
here).stats::predict() and utils::str()
explicitly (R CMD check NOTE); fixed an Rd link warning in an internal
comment.torch_basis_autograd() (validation) now uses one
backward pass instead of one per basis function; added a repeated-run
stability test after a NaN result was observed once on Windows.RUN_IN_RSTUDIO.R skips installation when torch is ready
and runs R CMD check directly, so Rtools is not
required.Rewrite of the earlier prototype, following phases 1-5 of the release plan.
G + 2k + 1 knots, pykan convention)
and Cox-de Boor basis in torch and in base R
(kan_bspline_basis(), kan_bspline_basis_r()),
analytic derivative (kan_bspline_deriv_r()).torch_clamp() domain restriction is replaced by
an explicit policy extrapolate = "linear" (default, C1
Taylor extension, non-zero gradient), "none" (pykan
behavior) or "clamp".right_closed).kan_validate_basis(),
kan_validate_derivative(),
kan_validate_gradient(), kan_validate_all():
comparisons with splines::splineDesign(), finite
differences and torch autograd.kan_forward_r(), kan_gradient_r(),
kan_fd_gradient_r()).inst/validation/pykan_reference.py.kan_fit(): Adam / L-BFGS, mini-batches, validation
split or data, early stopping with best-epoch restore, train-only
normalization of x and y, gradient clipping,
weight decay, seeds, checkpoints, warm start, base R backend.kan_benchmark() and kan_synthetic() for
multi-seed convergence studies.kan_update_grid()), grid
refinement (kan_refine()), pykan-style regularization
(kan_reg_control()), pruning (kan_prune()),
edge importance, edge-function, network, partial-dependence and history
plots.kan_save() / kan_load() store plain R data
(no torch pointers).