The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.
HMM$plot_2d(), for plotting a model component as a
surface over two covariates. HMM$plot() varies one
covariate and holds the rest fixed, which for a bivariate term such as
s(x, y) shows a single slice; this varies both and draws
the surface. It is basis-agnostic, so it works for a thin plate spline,
bs = "gp", a Gaussian field, or two covariates entering the
model separately. Cells far from any observation are left blank by
default (too_far, as in mgcv::plot.gam()), and
show = "ci" fills by the width of the confidence interval
instead of the estimate. A field over one covariate needs nothing new:
it is an mgcv smooth, and HMM$plot() already draws it like
any other.HMMTMB_AD_FRAMEWORK=CppAD before installing to fall back;
fitting a field then warns.s(x, y, bs = "spde", xt = list(mesh = mesh)) in two
dimensions, and s(x, bs = "spde") in one, where a mesh of
k quadratic B-splines on evenly spaced knots is built for
you over the range of the covariate widened by a fifth on each side.
k defaults to 15 rather than mgcv’s 10, because it is the
resolution of the field and not a cap on wiggliness: the range is
estimated and is what does the smoothing. The two parameters are a
marginal standard deviation and a range, reported by
lambda(). A model containing a field uses the banded
forward algorithm by default, with bw = 15. Needs the
fmesher package; nothing here needs INLA. See
?smooth.construct.spde.smooth.spec and
inst/examples/spde/spde_field.R.L
convention, so that a smooth may combine several penalty matrices
through fewer smoothing parameters, with
log(lambda) = L * theta. A smooth may also supply its own
starting values and parameter names. L is the identity and
the starting values are lambda = 1 for every basis mgcv
ships, so nothing changes for an existing model.bw argument of HMM$new(), with
HMM$update_bw() and HMM$check_bw(). It
truncates the conditioning of each log-likelihood contribution at a
fixed lag, which makes the Hessian with respect to latent variables
spread over time banded rather than dense, at about twice the cost per
evaluation and with an error that decays geometrically in the bandwidth.
Off by default, so the exact forward algorithm is unchanged.HMM$post_coeff() now samples a model with random
effects through a sparse Cholesky factorisation of the joint precision,
rather than inverting it first. The inverse is dense even when the
precision is not, so this is much cheaper and much smaller in memory for
a model with many random effects. Models without random effects are
unaffected.mgcv::gam() to a dummy response.
predict.gam(type = "lpmatrix") uses nothing that fitting
produces, so the unfitted setup is enough.make_matrices()
returned one determinant per formula rather than one per penalty, so the
likelihood read past the end of log_det_S, and the
smoothing parameters of such a model were biased.HMM$post_coeff() and HMM$confint()Forecast classHMM$fit_stan()HMM$pseudores()laplace = TRUE in
HMM$fit_stan()HMM$confint()delta0 when
initial_state = "stationary"par_alt() functions for pretty display of cat and
mvn parameters.HMM$suggest_initial()These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.