ModalNet 0.3.1

• The null model (intercept and dispersion only) is now always among the candidates compared by the EBIC. The automatic grid is built from the initial pseudo-response, so its first point could already contain active covariates (or a numerically negligible one), in which case the null model could not be selected. When that happens the null fit is computed directly and placed at the top of the path. Fits with a user-supplied lambda are unchanged.
• penalty.factor is rescaled to sum to p, as glmnet does internally. Before, with weights on another scale (adaptive Lasso) the line search compared a penalty different from the one minimized by the inner solver, rejected valid steps and the path reached maxit.
• At a grid point where the dispersion stays at the boundary of its range for ten consecutive iterations (exponential limit, mode near zero) the iteration now stops and the fit is flagged as not converged, instead of running to maxit.

ModalNet 0.3.0

• First CRAN release.
• Dashun Liu added as author.
• Documentation of every exported function and method now describes its return value and includes runnable examples; references carry DOIs.
• verbose = TRUE reports progress through message().
• print() labels alpha = 0 fits as ridge.

ModalNet 0.2.1

• Fixed the scale of the ridge part of the penalty for alpha < 1. glmnet standardizes the Gaussian response and therefore divides its ridge term by the standard deviation of the pseudo-response; the inner call now compensates, so the penalized objective is exactly lambda * sum(alpha * |b| + (1 - alpha) * b^2 / 2) on the standardized columns. Before the fix the line search rejected the glmnet step at some grid points and ridge paths reached maxit. Lasso fits are unchanged.

ModalNet 0.2.0

• Package renamed from ModalLasso to ModalNet; main function modal_net(). modal_lasso() remains as an alias.
• Elastic net and ridge penalties through alpha.