hmmTMB 2.0.0

• Add scaled chi-squared distribution “schisq”
• Add 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$plot() already draws it like any other.
• Compile with TMB’s TMBad framework rather than its CppAD default. A Gaussian field has to factorise a parameter-dependent sparse precision inside the likelihood, which CppAD records operation by operation onto the tape: on a 952-node two-dimensional mesh one gradient takes 382 s under CppAD against 1.1 s under TMBad. Set HMMTMB_AD_FRAMEWORK=CppAD before installing to fall back; fitting a field then warns.
• Add latent Gaussian fields, as a Matern SPDE smoother usable anywhere an mgcv smooth is: s(x, y, bs = "spde", xt = list(mesh = mesh)) in two dimensions, and s(x, bs = "spde") in one, where a mesh of k quadratic B-splines on evenly spaced knots is built for you over the range of the covariate widened by a fifth on each side. k defaults to 15 rather than mgcv’s 10, because it is the resolution of the field and not a cap on wiggliness: the range is estimated and is what does the smoothing. The two parameters are a marginal standard deviation and a range, reported by lambda(). A model containing a field uses the banded forward algorithm by default, with bw = 15. Needs the fmesher package; nothing here needs INLA. See ?smooth.construct.spde.smooth.spec and inst/examples/spde/spde_field.R.
• Generalise the smoothing penalty to mgcv’s L convention, so that a smooth may combine several penalty matrices through fewer smoothing parameters, with log(lambda) = L * theta. A smooth may also supply its own starting values and parameter names. L is the identity and the starting values are lambda = 1 for every basis mgcv ships, so nothing changes for an existing model.
• Add the banded forward algorithm of Fischer (2026), through the new bw argument of HMM$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$post_coeff() now samples a model with random effects through a sparse Cholesky factorisation of the joint precision, rather than inverting it first. The inverse is dense even when the precision is not, so this is much cheaper and much smaller in memory for a model with many random effects. Models without random effects are unaffected.
• Build prediction matrices without fitting a throwaway mgcv::gam() to a dummy response. predict.gam(type = "lpmatrix") uses nothing that fitting produces, so the unfitted setup is enough.
• Fix the log-determinants of the penalty matrices when one linear predictor contains several smooth terms. make_matrices() returned one determinant per formula rather than one per penalty, so the likelihood read past the end of log_det_S, and the smoothing parameters of such a model were biased.
• New vignette on (semi-)supervised learning
• Fix parameter counts for models with constraints
• Use safe Hessian inversion even for models without random effects in HMM$post_coeff() and HMM$confint()
• Forecasting through the Forecast class

hmmTMB 1.1.2

• Add hurdle negative binomial distribution
• Fix bug in HMM$fit_stan()
• Fix bug in HMM$pseudores()

hmmTMB 1.1.1

• Add error message for laplace = TRUE in HMM$fit_stan()
• Add reference to JSS paper

hmmTMB 1.1.0

• Add standard error to output of HMM$confint()
• Use ID-specific delta0 when initial_state = "stationary"
• Add par_alt() functions for pretty display of cat and mvn parameters.
• Improve mvn distribution (automatically detect dimension), and fix mvnorm for dim > 2, using a Cholesky decomposition to get unconstrained parameters of covariance matrix.
• Add zero-one-inflated beta distribution
• Replace optimx by nlminb for model fitting
• Allow empty models for simulation
• Use generalized determinant for penalty matrices
• Add nbinom(mean, shape) distribution
• Allow for user-specified arguments to be passed directly to mgcv::gam(); e.g., knots.
• Add built-in functions dvm, rvm, dwrpcauchy, rwrpcauchy, to remove dependence on CircStats (as requested by CRAN).

hmmTMB 1.0.2

• Fix bug for categorical distribution
• Fix bug when using tibble input data
• Fix bugs for models with shared parameters
• Add HMM$suggest_initial()