BIE::ParallelChains Class Reference

Implements parallel temperering or the parallel hierarchical sampler. More...

#include <ParallelChains.h>

Inheritance diagram for BIE::ParallelChains:

Collaboration diagram for BIE::ParallelChains:

List of all members.

Global variables

enum Algorithm { standard, hierarchical }

Algorithm types. More...

enum Control { parallel, serial }

Parallelization type. More...

enum Initial { user_supplied, prior_sampled }

Initialization type. More...

static Algorithm algo

Current algorithm.

static Control cntrl

Current control.

static Initial initial

Initialization method.

static double swapprob

Swap probability.

static double apow

Change temperature spacing (apow=1 is geometric spacing).

static double tpow

Scale factor $T^{tpow}$ for proposal function.

static int minmc

Minimum number of temperature states.

static int state_iter

Number of iterations allowed for finding a good initial state.

vector< unsigned short > fail

Status vector for sampling new states.

void SampleNewState (int k)

Attempt to find a good state from the prior distribution for chain k.

void SampleNewStateParallel (int k)

Sample a new state for the kth chain (parallel version).

void SampleNewStateSerial (int k)

Sample a new state for the kth chain (serial version).

void NewStateSerialInit ()

Prepare for chain initialization.

void NewStateSerialStatus ()

Check the status of the initialization.

Public Member Functions

virtual ~ParallelChains ()

Destructor.

virtual void Initialize ()

Initialize all the chains.

virtual void Reinitialize (MHWidth *width, MixturePrior *mp)

Reinitialize with new class instances.

ParallelChains (int max, int ndim, int minmc_p, double maxT, MHWidth *width, BaseDataTree *d, Model *m, Integration *i, Converge *c, MixturePrior *mp, LikelihoodComputation *l, MCAlgorithm *mca)

Constructor.

ParallelChains (int max, int ndim, int minmc_p, double maxT, MHWidth *width, BaseDataTree *d, Model *m, Integration *i, Converge *c, MixturePrior *mp, LikelihoodComputation *l, MCAlgorithm *mca, Simulation *last)

void NewState (int Mcur, vector< double > &w, vector< vector< double > > &p)

Initialize new state (override from Simulation).

void NewState (Chain &ch)

void NewState (State &s)

void SetAlgorithm (int)

Set global parameters.

void SetControl (int)

Choose parallelization control (parallel=0, serial=1).

void SetInitial (int)

Choose initialization type (user_supplied=0, prior_sampled=1).

void SetTempExp (double a, double t)

Choose temperature scaling parameters.

void SetNewWidths (vector< double > *wid)

Set chains widths to new values.

void SetNewWidths (string wid)

Set chains widths to new values (file name input).

Members which report and diagnose current state

double GetValue ()

Return current posterior probability value (override).

double GetPrior ()

Return current prior value (override).

double GetLikelihood ()

Return current likelihood probability value.

State GetState ()

Return current state (override).

virtual void ReportState ()

Print passed state in ascii with labels (override).

void PrintState ()

Print current state (override).

virtual void PrintStepDiagnostic ()

override of Simulation method to provide diagnostics specific to temperedsimulation

virtual void PrintStateDiagnostic ()

override of Simulation method to provide state specific diagnostics.

vector< double > GetMixstat (void)

Return vector with fraction of swaps for each chain.

vector< double > GetSweepstat (void)

Return vector with fraction of acceptences for each chain.

void NewNumber (int)

Specify the number of desired chains.

void NewSwapProb (double p)

Change the swap probabability.

void EnableLogging ()

Enable caching per chain.

void AdditionalInfo ()

Print out simulation specific diagnostics.

Protected Member Functions

virtual void MCMethod ()

Execute a Markov Chain step.

void CreateChains ()

Create the chain structures.

void ResetChainStats ()

For gathering diagnostic data.

void getChainCounts ()

Gathers in-chain rates across all processes.

Protected Attributes

bool caching

True if chains states are logged to files.

unsigned ncount

Number of attempts so far.

Uniform * unit

Unit interval variates.

bool chains_initialized

True if state values have been computed for at least the first time.

vector< double > user_widths

User-defined widths for MH sampling.

double MaxT

Config parameters.

vector< unsigned char > update_flag

State variables.

vector< unsigned char > update_flag1

List of chains on an individual node that must sychronized.

vector< unsigned long > stat0

Number of in-chain update attempts.

vector< unsigned long > stat1

Number of in-chain acceptances per chain.

vector< unsigned long > nstat0

Number of swap attempts per chain.

vector< unsigned long > nstat1

Number of swap acceptances per chain.

vector< unsigned long > mstat0

Number of standard MCMC updates attempted per chain.

vector< unsigned long > mstat1

Number of standard MCMC updates accepted per chain.

int Ntot

Maximum dimension of state vector.

int swapnum

Convergence check.

int swaptry

The number of attempted swaps.

Friends

class boost::serialization::access

Detailed Description

Implements parallel temperering or the parallel hierarchical sampler.

: Both algorithms use an ensemble of Markov chains, run in parallel, each at a successively higher temperature. A temperature of is the initial cold chain.

The total number of chains is given by the maximum temperature, MaxT, for a state vector of dimension Ntot:

$//! Mchains = 1 + \log(MaxT)*\sqrt(Ntot) //!$

Temperature spacing

Temperature for a chain is chosen as follows:

$//! T_j = T_{max}^{(j/(Mchain-1))^a}, j\in[0,Mchain-1] //!$

is logarithmic (e.g pure geometric spacing). This is the default.
weighted towards values of smaller than geometric
weighted towards values of larger than geometric
gives values of approximately linear from 1 to $T_{max}$
gives values of strongly weighted toward $T_{max}$

The second exponent, $t$ is used to scale the proposal width by a power of the temperature. Gaussian scaling, the default is $t=1/2$ .

The member function SetTempExp may be used to rest either or both of these parameters.

Parallel tempering (standard)

At each step, with probability swapprob, one proposes to swap the states of two adjecent temperature chains. With probability 1 - swapprob, all chains are updated at their assigned tempratures.

Parallel hierarchical sampler: At each step, one chain is proposed to be swapped with the cold chain. The remaining Mchain-2 schains are updated at their assigned tempratures.

Note:

Chain initialization The fiducial, cold chain is initialized from the user-supplied initial state file. The initialization of the warm state can be done in two way depending on the static variable initial.

If initial = user_supplied, all chains are assigned the user-supplied initial value.
If initial = prior_sampled, all chains but the fiducial chain have initial values sampled from the prior distribution.

Control methods

The parallel likelihood classes should be used with parallel control only. The parallel likelihood classes LikelihoodComputationMPI and LikelihoodComputationMPITP assume that each process will enter the likelihood computation code.
A serial likelihood class should be used with serial control only. A serial likelihood computation is only entered by the process which calls it.

Diagnostic output A call to the EnableLogging member function will cause the state of each individual chain to be logged to a separate file. The chain files will have names of the form "nametag.chainlog.k" where nametag is the global nametag variable and k is the chain number. The first five lines record the following values for each chain: the number of components in the mixture, rank of the parameter vector, the inverse temperature $\beta=1/T$ , the Metropolis-Hastings width factor, and the weighting factor.

References

: Geyer, C. J. 1991, "Markov chain Monte Carlo maximum likelihood." In Keramidas, E. M., editor, Computing Science and Statistics, Proceedings of the 23rd Symposium on the Interface, pages 156-163, Seattle, WA. Interface Foundation of North America.

: Hukushima and Nemoto 1996, J. Phys. Soc. Japan, 65, 1604-1620.

: Rigat, F. 2006, "MCMC inference using parallel hierarchical sampling," preprint.

Member Enumeration Documentation

enum BIE::ParallelChains::Algorithm

Algorithm types.

Enumerator:

standard	standard parallel chains algorithm.
hierarchical	hierarchical parallel chains algorithm.

enum BIE::ParallelChains::Control

Parallelization type.

Enumerator:

parallel	root controls all chain updates.
serial	chains are run at each node.

enum BIE::ParallelChains::Initial

Initialization type.

Enumerator:

user_supplied	initial state specified by user.
prior_sampled	initial state selected from prior.

Member Function Documentation

void BIE::ParallelChains::SetAlgorithm ( int )

Set global parameters.

Choose between parallel tempering and hierachical tempering (standard=0, hierarchical=1)

void BIE::ParallelChains::SetNewWidths ( vector< double > * wid )

Set chains widths to new values.

Set chains widths to new values (vector input)

virtual void BIE::ParallelChains::PrintStateDiagnostic ( ) [virtual]

override of Simulation method to provide state specific diagnostics.

Here, the method provides mixing statistics.

Reimplemented from BIE::Simulation.

void BIE::ParallelChains::SampleNewState ( int k ) [protected]

Attempt to find a good state from the prior distribution for chain k.

Sample a new state for the kth chain (will select one of the two below)

void BIE::ParallelChains::ResetChainStats ( ) [protected]

For gathering diagnostic data.

Resets the counters

Member Data Documentation

double BIE::ParallelChains::MaxT [protected]

Config parameters.

Temperature for tempering ladder

vector<unsigned char> BIE::ParallelChains::update_flag [protected]

State variables.

List of chains that must be globally sychronized

int BIE::ParallelChains::swapnum [protected]

Convergence check.

The number swaps

The documentation for this class was generated from the following file:

/home/weinberg/src/BIE/include/ParallelChains.h