Estimate Parameters of a drift_dm Model via Classical Optimization

estimate_classical() estimates the parameters of a drift_dm model by minimizing the model's cost function (e.g., RMSE or negative log-likelihood) using classical (non-Bayesian) optimization routines.

Available optimizers include:

Nelder-Mead (bounded or unbounded): "Nelder-Mead", "nmkb" (via stats::optim() and dfoptim::nmkb(), respectively)
BFGS and L-BFGS-B (via stats::optim())
Differential Evolution (via DEoptim::DEoptim())

Usage

estimate_classical(
  drift_dm_obj,
  optimizer,
  start_vals = NULL,
  return_runs = NULL,
  lower = NULL,
  upper = NULL,
  verbose = NULL,
  de_n_cores = 1,
  control = list(),
  round_digits = NULL,
  seed = NULL
)

Arguments

drift_dm_obj

an object inheriting from drift_dm.

optimizer

a character string specifying the optimizer to use. Must be one of "Nelder-Mead", "nmkb", "BFGS", "L-BFGS-B", or "DEoptim".

start_vals

a set of starting values. Must be compatible with get_parameters_smart(). If start_vals is not NULL, the function tries to set the provided parameter values to the model, using those values as starting values for the optimization routine. Special case: If start_vals is a data.frame, the function is recursively called for each row of start_vals, providing a handy way to run an optimization routine with different starting values. Default is NULL, which implies that the current model parameters are used as starting values.

return_runs

a single logical. Only relevant when start_vals is a data.frame and the optimization routine is called multiple times with different starting values. If FALSE, the best-fitting model is returned. If TRUE, a list is returned, containing the best-fitting model, all cost values across runs, and all estimated model parameters across runs.

lower, upper

bounds on the parameters to be estimated. Can be numeric vectors, named vectors, or flexible lists (see Details).

verbose

an integer (0, 1, or 2). Controls the amount of printed output.

0 = silent
1 = starting/exiting messages
2 = all parameters and the cost value per iteration

de_n_cores

an integer > 0. Number of CPU cores to use for DEoptim.

control

a named list of control parameters passed to the chosen optimizer.

round_digits

an integer. Number of digits to round cost values in printed output. If NULL, defaults to drift_dm_default_rounding().

seed

a seed, to make the results of DEoptim reproducible.

Value

The updated drift_dm_obj, with optimized parameters.

Details

Search space specification

lower and upper can be specified flexibly:

As unnamed numeric vectors (not recommended unless you're sure of the parameter order)
As named numeric vectors matching the parameter names of the model
As lists with a default_values entry (plus optional condition-specific entries)

This design mirrors the structure used in simulate_data.drift_dm().

Optimization details

Some optimizers (i.e., "nmkb", "L-BFGS-B", "DEoptim") require both lower and upper bounds.

Differential Evolution (DEoptim) supports parallelization across cores via de_n_cores. If de_n_cores > 1, a parallel cluster is created and automatically closed after optimization.

The cost function being minimized depends on the cost_function() of the model.

During optimization, failed model evaluations yield a very high cost value (i.e., .Machine$double.xmax). In some cases, this ensures that the optimization doesn't crash, though, this is not guaranteed.