twDEMCSACont: twDEMCSACont

Description Usage Arguments Details Author(s)

View source: R/sa.R

Description

continuing simulated annealing DEMC based on previous result

Usage

1
2
3
4
5
6
7
8
twDEMCSACont(mc, nGen = 512, nObs, ..., m0 = calcM0twDEMC(getNParms(mc), 
    getNChainsPop(mc)), controlTwDEMC = list(), debugSequential = FALSE, 
    restartFilename = NULL, ctrlBatch = list(nSamplesBatch = m0 * 
        8, pThinPast = 0.5), ctrlT = list(TFix = numeric(0), 
        TMax = numeric(0),
 TDecProp = 0.9), ctrlConvergence = list(maxRelTChangeCrit = 0.05, 
        minTBase0 = 1e-04, maxLogDenDriftCrit = 0.3, gelmanCrit = 1.4, 
        critSpecVarRatio = 20, dumpfileBasename = NULL, maxThin = 64))

Arguments

mc

result of twDEMCBlock

nGen

overall number of generations to add (within one batch provide ctrlBatch$nGenBatch)

nObs

integer vector (nResComp) specifying the number of observations for each result component

...

further argument to twDEMCBlockInt

m0

minimum number of samples in step for extending runs

controlTwDEMC

list argument to twDEMCBlockInt containing entry thin

debugSequential

set to TRUE to avoid parallel execution, good for debugging

restartFilename

filename to write intermediate results to

ctrlBatch

list of arguments controlling batch executions

nSamplesBatch

number of samples for one call to twDEMCStep (multiplied by thin to get generations), defaults to 8*m0 << default: set in a way that on average each population (assuming half are significant) is appended by 2*m0 samples

pThinPast

in each batch thin the past to given fraction before appending new results

ctrlT

list of arguments controlling Temperature decrease

TFix

numeric vector (nResComp) specifying a finite value for components with fixed Temperatue, and a non-finite for others

TMax

numeric vector (nResComp) specifying a maximum temperature for result components.

TDecProp

proportion of Temperature decrease: below one to diminish risk of decreasing Temperature too fast (below what is supported by other data streams)

ctrlConvergence

list or arguments controlling check for convergence

maxRelTChangeCrit

if Temperature of the components changes less than specified value, the algorithm can finish

minTBase0

if Temperature-1 gets below this temperature, the algorithm can finish

maxLogDenDriftCrit

if difference between mean logDensity of first and fourth quartile of the sample is less than this value, we do not need further batches because of drift in logDensity

gelmanCrit

do not change Temperature, if variance between chains is too high, i.e. Gelman Diag is above this value

critSpecVarRatio

if proprotion of spectral Density to Variation is higher than this value, signal problems and resort to subspaces

dumpfileBasename

scalar string: filename to dump stack before stopping. May set to "recover"

maxThin

scalar positive integer: maximum thinning interval. If not 0 then thinning is increased on too high spectral density

Details

diagnostics

in resulting twDEMC an entry diagnostics is returned. Its a list with diagnostics after each batch. Its components have been calculated on the last halv of the chains.

T0

Temperature (variance inflation factor) scaled for one observation minus one

maxGelmanDiagPops

maximum of within population gelman diagnostics

gelmanDiag

gelman diag calculated between populations

maxSpecVarRatio

ratio of spectral density to sample Variance (max over all parameters). This is a measure that increases with autocorrelation.

Adaptive thinning interval (maxThin)

If spectral variance is larger than its critical ratio ctrlConvergence$critSpecVarRatio then thinning interval needs to be increased. If doubling the thinning interval is not larger than ctrlConvergence$maxThin then a new batch with higher thinning is attempted. Else the SADEMC quits with reporting an error with return values component failureMsg.

Author(s)

Thomas Wutzler


twDEMC documentation built on May 31, 2017, 3:44 a.m.