NEWS
thangldKANr 0.1.0 (2026-11-20)
- First release.
thangld_kan(widths, basis = ...): Kolmogorov-Arnold networks with
Chebyshev, RBF (FastKAN), Legendre, wavelet, rational, hybrid, Jacobi and
ReLU-KAN edges; per-layer bases; basis-specific arguments in ... or
options.
thangld_graph_kan(): graph KANs for transductive node regression
(aggregate with the normalized adjacency, optional self branch), with
fit, predict, save/load and thangld_graph_kan_validate().
thangld_temporal_kan(): lag-window KANs for one-step forecasts, with
fit, predict, rolling-origin evaluation and
thangld_temporal_kan_validate().
- Both accept every edge basis.
register_kan_basis(), unregister_kan_basis(), kan_bases(),
kan_layer_bases(): add edge bases from other packages.
thangld_kan_validate(backend, basis, mixed): numerical checks of every
basis and of a network mixing bases.
inst/validation/<name>: comparisons with the Python reference of each
basis, of the graph KAN and of the lag-window KAN at pinned commits.