add.simulation: Create or augment a list of simulated distributions of...

In Infusion: Inference Using Simulation

Create or augment a list of simulated distributions of summary statistics

Description

add_simulation is suitable for the primitive Infusion workflow; otherwise, it is cleaer to call add_reftable directly. add_simulation creates or augments a list of simulated distributions of summary statistics, and formats the results appropriately for further use. Alternatively, if the simulation function cannot be called directly by the R code, simulated distributions can be added using the newsimuls argument, using a simple format (see onedistrib in the Examples). Finally, a generic data frame of simulations can be reformatted as a reference table by using only the simulations argument.

Depending on the arguments, parallel or serial computation is performed. When parallelization is implied, by default a “socket” cluster, available on all operating systems. Special care is then needed to ensure that all required packages are loaded in the called processes, and that all required variables and functions are passed therein: check the packages and env arguments. For socket clusters, foreach or pbapply is called depending whether the doSNOW package is attached (doSNOW allows more efficient load balancing than pbapply).

Usage

add_simulation(simulations=NULL, Simulate, parsTable=par.grid, par.grid=NULL, 
               nRealizations=Infusion.getOption("nRealizations"),
               newsimuls=NULL, verbose=interactive(), nb_cores=NULL, 
               packages=NULL, env=NULL, control.Simulate=NULL,
               cluster_args=list(), cl_seed=NULL, ...) 

Arguments

simulations	A list of matrices each representing a simulated distribution for given parameters in a format consistent with the return format of add_simulation.
nRealizations	The number of simulated samples of summary statistics, for each empirical distribution (each row of par.grid).
Simulate	An *R* function, or the name (as a character string) of an *R* function used to generate empirical distributions of summary statistics. When an external simulation program is called, Simulate must therefore be an R function wrapping the call to the external program. The Simulate function must have one argument for each element of the parameter vector (i.e. of each row of par.grid). It must return a vector of summary statistics with named vector members; or a single matrix of nRealizations simulations, in which case its rows and row names must represent the summary statistics, it should have nRealizations columns, and nRealizations should be named integer of the form “c(as_one=.)” (see Examples).
parsTable, par.grid	A data frame of which each line is the vector of parameters needed by Simulate for each simulation of the data-generating process. par.grid is an alias for parsTable; the latter argument may be preferred in order not to suggest that the parameter values should form a regular grid.
newsimuls	If the function used to generate empirical distributions cannot be called by R, then newsimuls can be used to provide these distributions. See Details for the structure of this argument.
nb_cores	Number of cores for parallel simulation; NULL or integer value, acting as a shortcut for cluster_args$spec. The effect is complicated: see Details.
cluster_args	A list of arguments, passed to makeCluster. May contain a non-null spec element, in which case the distinct nb_cores argument and the global Infusion option nb_cores are ignored. A typical usage would thus be control_args=list(spec=<number of 'children'>). Additional elements outfile="log.txt" may be useful to collect output from the nodes, and type="FORK" may be used to force a fork cluster on linux(-alikes) (otherwise a socket cluster is set up as this is the default effect of parallel::makeCluster).
verbose	Whether to print some information or not.
...	Arguments passed to add_reftable (and possibly beyond, to the simulation function: see nsim argument of myrnorm_tab() in the Examples. These arguments should be constant through all the simulation workflow.
control.Simulate	A list, used as an exclusive alternative to “...” to pass additional arguments to Simulate, beyond the parameter vector. The list must contain the same elements as would go in the “...”.
packages	For parallel evaluation: Names of additional libraries to be loaded on the cores, necessary for Simulate evaluation.
env	For parallel evaluation: an environment containing additional objects to be exported on the cores, necessary for Simulate evaluation.
cl_seed	Integer, or NULL. Providing the seed was conceived to allow repeatable results at least for given parallelization settings, if not identical results across different parallelization contexts. However, this functionality may have been been lost as the code was adapted for the up-to-date workflow using add_reftable.

Details

The newsimuls argument should have the same structure as the return value of the function itself, except that newsimuls may include only a subset of the attributes returned by the function. newsimuls should thus be list of matrices, each with a par attribute (see Examples). Rows of each matrix stand for simulation replicates and columns stand for the different summary statistics.

When nRealizations>1L, if nb_cores is unnamed or has name "replic" and if the simulation function does not return a single table for all replicates (thus, if nRealizations is not a named integer of the form “c(as_one=.)”, parallelisation is over the different samples for each parameter value (and the seed of the random number generator is not controlled in a parallel context). For any other explicit name (e.g., nb_cores=c(foo=7)), or if nRealizations is a named integer of the form “c(as_one=.)”, parallelisation is over the parameter values (the rows of par.grid). In all cases, the progress bar is over parameter values. See Details in Infusion.options for the subtle way these different cases are distinguished in the progress bar.

Using a FORK cluster with nRealizations>1 is warned as unreliable: in particular, anyone trying this combination should check whether other desired controls, such as random generator seed, or progress bar are effective.

Value

If nRealizations>1L, the return value is an object of class EDFlist, which is a list-with-attributes of matrices-with-attribute. Each matrix contains a simulated distribution of summary statistics for given parameters, and the "par" attribute is a 1-row data.frame of parameters. If Simulate is used, this must give all the parameters to be estimated; otherwise it must at least include all variable parameters in this or later simulations to be appended to the simulation list.

The value has the following attributes: LOWER and UPPER which are each a vector of per-parameter minima and maxima deduced from any newsimuls argument, and optionally any of the arguments Simulate, control.Simulate, packages, env, par.grid and simulations (all corresponding to input arguments when provided, except that the actual Simulate function is returned even if it was input as a name).

If nRealizations=1 add_reftable is called: see its distinct return value.

Examples

### Examples using init_grid and add_simulation, for primitive workflow
### Use init_reftable and add_reftable for the up-to-date workflow

# example of building a list of simulations from scratch:
myrnorm <- function(mu,s2,sample.size) {
  s <- rnorm(n=sample.size,mean=mu,sd=sqrt(s2))
  return(c(mean=mean(s),var=var(s)))
}
set.seed(123)
onedistrib <- t(replicate(100,myrnorm(1,1,10))) # toy example of simulated distribution
attr(onedistrib,"par") <- c(mu=1,sigma=1,sample.size=10) ## important!
simuls <- add_simulation(NULL, Simulate="myrnorm", nRealizations=500,
                         newsimuls=list("example"=onedistrib))

# standard use: smulation over a grid of parameter values
parsp <- init_grid(lower=c(mu=2.8,s2=0.2,sample.size=40),
                   upper=c(mu=5.2,s2=3,sample.size=40))
simuls <- add_simulation(NULL, Simulate="myrnorm", nRealizations=500,
                         par.grid = parsp[1:7,])
                         
## Not run:  # example continued: parallel versions of the same
# Slow computations, notably because cluster setup is slow.

#    ... parallel over replicates, serial over par.grid rows
# => cl_seed has no effect and can be ignored
simuls <- add_simulation(NULL, Simulate="myrnorm", nRealizations=500,
                         par.grid = parsp[1:7,], nb_cores=7)
#                         
#    ... parallel over 'par.grid' rows => cl_seed is effective
simuls <- add_simulation(NULL, Simulate="myrnorm", nRealizations=500,
                         cl_seed=123, # for repeatable results
                         par.grid = parsp[1:7,], nb_cores=c(foo=7))

## End(Not run)
                     
####### Example where a single 'Simulate' returns all replicates:

myrnorm_tab <- function(mu,s2,sample.size, nsim) {
  ## By default, Infusion.getOption('nRealizations') would fail on nodes!
  replicate(nsim, 
            myrnorm(mu=mu,s2=s2,sample.size=sample.size)) 
}

parsp <- init_grid(lower=c(mu=2.8,s2=0.2,sample.size=40),
                   upper=c(mu=5.2,s2=3,sample.size=40))

# 'as_one' syntax for 'Simulate' function returning a simulation table: 
simuls <- add_simulation(NULL, Simulate="myrnorm_tab",
              nRealizations=c(as_one=500),
              nsim=500, # myrnorm_tab() argument, part of the 'dots'
              par.grid=parsp)

## Not run:  # example continued: parallel versions of the same.
# Slow cluster setup again
simuls <- add_simulation(NULL,Simulate="myrnorm_tab",par.grid=parsp,
              nb_cores=7L,
              nRealizations=c(as_one=500),
              nsim=500, # myrnorm_tab() argument again
              cl_seed=123, # for repeatable results
              # need to export other variables used by *myrnorm_tab* to the nodes:
              env=list2env(list(myrnorm=myrnorm)))

## End(Not run)

## see main documentation page for the package for other typical usage

Infusion documentation built on May 3, 2023, 5:10 p.m.