Detailed control for ecostate structure — ecostate

Define a list of control parameters.

Usage

ecostate_control(
  nlminb_loops = 1,
  newton_loops = 0,
  eval.max = 1000,
  iter.max = 1000,
  getsd = TRUE,
  silent = getOption("ecostate.silent", TRUE),
  trace = getOption("ecostate.trace", 0),
  verbose = getOption("ecostate.verbose", FALSE),
  profile = c("logF_ti", "log_winf_z", "s50_z", "srate_z"),
  random = c("epsilon_ti", "alpha_ti", "nu_ti", "phi_tg2"),
  tmb_par = NULL,
  map = NULL,
  getJointPrecision = FALSE,
  integration_method = c("ABM", "RK4", "ode23", "rk4", "lsoda"),
  process_error = c("epsilon", "alpha"),
  n_steps = 10,
  F_type = c("integrated", "averaged"),
  derived_quantities = c("h_g2", "B_ti", "B0_i"),
  scale_solver = c("joint", "simple"),
  inverse_method = c("Standard", "Penrose_moore"),
  tmbad.sparse_hessian_compress = 1,
  start_tau = 0.001
)

Arguments

nlminb_loops: Integer number of times to call stats::nlminb().
newton_loops: Integer number of Newton steps to do after running stats::nlminb().
eval.max: Maximum number of evaluations of the objective function allowed. Passed to control in stats::nlminb().
iter.max: Maximum number of iterations allowed. Passed to control in stats::nlminb().
getsd: Boolean indicating whether to call TMB::sdreport()
silent: Disable terminal output for inner optimizer?
trace: Parameter values are printed every trace iteration for the outer optimizer. Passed to control in stats::nlminb().
verbose: Output additional messages about model steps during fitting?
profile: parameters that are profiled across, passed to MakeADFun
random: parameters that are treated as random effects, passed to MakeADFun
tmb_par: list of parameters for starting values, with shape identical to tinyVAST(...)$internal$parlist
map: list of mapping values, passed to RTMB::MakeADFun
getJointPrecision: whether to get the joint precision matrix. Passed to sdreport.
integration_method: What numerical integration method to use. "ABM" uses a native-R versions of Adam-Bashford, "RK4" uses a native-R version of Runge-Kutta-4, and "ode23" uses a native-R version of adaptive Runge-Kutta-23, where all are adapted from pracma functions. "rk4" and lsoda use those methods from deSolve::ode as implemented by RTMBode::ode
process_error: Whether to include process error as a continuous rate (i.e., an "innovation" parameterization, process_error="epsilon") or as a discrete difference between expected and predicted biomass (i.e., a "state-space" parameterization), process_error="alpha"The former is more interpretable, whereas the latter is much more computationally efficient.
n_steps: number of steps used in the ODE solver for biomass dynamics
F_type: whether to integrate catches along with biomass ("integrated") or calculate catches from the Baranov catch equation applied to average biomass ("averaged")
derived_quantities: character-vector listing objects to ADREPORT
scale_solver: Whether to solve for ecotrophic efficiency EE given biomass B (scale_solver="simple") or solve for a combination of EE and B values
inverse_method: whether to use pseudoinverse or standard inverse
tmbad.sparse_hessian_compress: passed to TMB::config(), and enabling an experimental feature to save memory when first computing the inner Hessian matrix. Using tmbad.sparse_hessian_compress=1 seems to have no effect on the MLE (although users should probably confirm this), and hugely reduces memory use in both small and large models. Using tmbad.sparse_hessian_compress=1 seems to hugely speed up the model-fitting with a large model but results in a small decrease in speed for model-fitting with a small model.
start_tau: Starting value for the standard deviation of process errors