pauwels2014 (version 1.0)

BFGS_special: An implementation of BFGS method for posterior maximization.

Description

Gradients are computed using finite differences.

Usage

BFGS_special(init, knobj, fun_like, verbose = FALSE)

Arguments

init

An initial value of the parameter to be optimized.

knobj

A knowledge list. See knobjs.

fun_like

A function to compute posterior value. See eval_log_like_knobj

verbose

Print progresses of the local search?

Value

• A list with the following entries:
• theta
  The local optimum found by the method.
• fail
  A boolean representing wither the local search failed or not due to numerical problems.

Details

The step size are chosen using Armijo's rule. Special checks are performed to avoid numerical instabilities in the differential equation solver.

Examples

data(experiment_list1)
data(observables)

## Generate the knowledge object with correct parameter value
knobj <- generate_our_knowledge(transform_params)

## Initialize with some data
knobj$datas[[1]] <- list(
 manip = experiment_list1$nothing,
 data = add_noise(
  simulate_experiment(knobj$global_parameters$true_params_T, knobj, experiment_list1$nothing)[
   knobj$global_parameters$tspan %in% observables[["mrnaLow"]]$reso, 
   observables[["mrnaLow"]]$obs
  ]
 )
)
knobj$experiments <- paste("nothing", "mrnaLow")

theta <- rep( 50, length(knobj$global_parameters$param_names) )
names(theta) <- knobj$global_parameters$param_names

## Only perform 5 iterations
knobj$global_parameters$max_it <- 5

temp <- BFGS_special(theta, knobj, eval_log_like_knobj)
temp$theta