inference.cpp File Reference

Model inference tools. More...

#include <unistd.h>
#include <sstream>
#include <map>
#include <iostream>
#include <string>
#include <math.h>
#include "func.h"
#include "post.h"
#include "mrv.h"
#include "inference.h"
#include "mcmc.h"
#include "amcmc.h"
#include "tools.h"
#include "arrayio.h"
#include "arraytools.h"

Functions
void	infer_model (Array1D< Array2D< double >(*)(Array2D< double > &, Array2D< double > &, Array2D< double > &, void *) > forwardFuncs, void *funcInfo, string likType, string priorType, double priora, double priorb, Array2D< double > &xdata, Array1D< Array1D< double > > &ydata, Array2D< double > &xgrid, int dataNoiseInference, Array1D< double > &datanoise_array, int pdim, int order, Array1D< int > &rndInd, Array2D< double > &fixindnom, string pdfType, string pcType, int seed, int nmcmc, double mcmcgamma, bool optimflag, Array1D< double > &chstart, Array1D< double > &chsig, double likParam, int likParam_int, Array2D< double > &pgrid, Array2D< double > &pchain, int nburn, int nstep, Array1D< double > &mapparam, Array1D< double > &datavar_map, Array1D< double > &pmean_map, Array1D< double > &pvar_map, Array1D< double > &fmean_map, Array1D< double > &fvar_map, Array1D< double > &postave_datavar, Array1D< double > &p_postave_mean, Array1D< double > &p_postave_var, Array1D< double > &p_postvar_mean, Array2D< double > &f_postsam_mean, Array1D< double > &f_postave_mean, Array1D< double > &f_postave_var, Array1D< double > &f_postvar_mean, Array2D< double > &paramPCcfs)
	Main function for inferring model parameters.
double	LogPosterior (Array1D< double > &m, void *mypost_void)
	Log-posterior function given a vector of parameters (chain state) and a void* set of auxiliary variables.

Detailed Description

Model inference tools.

Author

K. Sargsyan 2016 -

Function Documentation

infer_model()

void infer_model	(	Array1D< Array2D< double >(*	forwardFuncs )(Array2D< double > &, Array2D< double > &, Array2D< double > &, void *) >,
		void *	funcInfo,
		string	likType,
		string	priorType,
		double	priora,
		double	priorb,
		Array2D< double > &	xdata,
		Array1D< Array1D< double > > &	ydata,
		Array2D< double > &	xgrid,
		int	dataNoiseInference,
		Array1D< double > &	datanoise_array,
		int	pdim,
		int	order,
		Array1D< int > &	rndInd,
		Array2D< double > &	fixIndNom,
		string	pdfType,
		string	pcType,
		int	seed,
		int	nmcmc,
		double	mcmcgamma,
		bool	optimflag,
		Array1D< double > &	chstart,
		Array1D< double > &	chsig,
		double	likParam,
		int	likParam_int,
		Array2D< double > &	pgrid,
		Array2D< double > &	pchain,
		int	nburn,
		int	nstep,
		Array1D< double > &	mapparam,
		Array1D< double > &	datavar_map,
		Array1D< double > &	pmean_map,
		Array1D< double > &	pvar_map,
		Array1D< double > &	fmean_map,
		Array1D< double > &	fvar_map,
		Array1D< double > &	postave_datavar,
		Array1D< double > &	p_postave_mean,
		Array1D< double > &	p_postave_var,
		Array1D< double > &	p_postvar_mean,
		Array2D< double > &	f_postsam_mean,
		Array1D< double > &	f_postave_mean,
		Array1D< double > &	f_postave_var,
		Array1D< double > &	f_postvar_mean,
		Array2D< double > &	paramPCcfs )

Main function for inferring model parameters.

Note

This is written in a fortran style, i.e. some arguments are inputs, and the rest are output

Parameters

[in]	*forwardFuncs	: an array of y=f(p,x) functions that take np-by-pdim and nx-by-xdim input arrays and returns an np-by-nx output, see tools/func.h for several examples
[in]	funcinfo	: auxiliary function-specific information (can be 0)
[in]	likType	: likelihood type: options are 'full', 'marg', 'mvn', 'gausmarg', 'abc', 'abcm', 'classical', 'koh' (see UQTk Manual)
[in]	priorType	: prior type: options are 'uniform', 'normal', 'inverse', etc ... (see UQTk Manual)
[in]	priora,priorb	: prior parameters (todo: need to be made dimension-specific) for uniform, it is the range, for normal it is the moments
[in]	xdata	: x-values of data, nx-by-xdim
[in]	ydata	: y-values of data, nx-by-neach
[in]	xgrid	: x-values where predictive moments are computed after the inference : can be the same as xdata
[in]	dataNoiseInference	: indicator, data noise stdev is fixed(0), inferred (1), or log-inferred (2)
[in]	datanoise_array	: data noise stdev array, if fixed, otherwise merely an MCMC starting point
[in]	pdim	: model parameter dimensionality
[in]	order	: order of output PC that is computed via NISP in the likelihood
[in]	rndInd	: array of indices of parameters to be randomized
[in]	fixIndNom	: array of indices and nominal values of parameters to be fixed
[in]	pdfType	: type of PDF PC parameterization, options are 'pct','pci' and 'full' (see UQTk Manual)
[in]	pcType	: type of PC for the PDF parameterization, options are all common PC types, e.g. 'HG','LU'
[in]	seed	: integer seed for MCMC
[in]	nmcmc	: number of MCMC steps to follow optimization; if 0, then only optimization is performed if optimflag is True
[in]	mcmcgamma	: gamma (scaling) parameter for adaptive MCMC
[in]	optimflag	: indicates if optimization is prepended to MCMC
[in]	chstart	: initial chain state
[in]	chsig	: initial non-adaptive, dimensionwise proposal jump size
[in]	likParam,likParam_int	: likelihood parameters (currently, only needed for KDE-based likelihoods, 'full' and 'marg', to pass the KDE bandwidth and number of samples)
[in]	pgrid	: parameter grid, if requested, to compute exact posterior (can be empty)
[in]	nburn	: burn-in for MCMC to write to pchain
[in]	nstep	: thinning of MCMC to write to pchain
[in,out]	pchain	: thinned chain file with nburn and nstep applied, to be used for postprocessing : if given as non-empty array, the MCMC is skipped, and only postprocessing is performed
[out]	mapparam	: MAP parameters
[out]	datavar_map	: MAP value of data variances
[out]	pmean_map,pvar_map	: MAP values of parameter means and variances
[out]	fmean_map,fvar_map	: MAP values of function predition (at xgrid) means and variances
[out]	postave_datavar	: posterior average of data variances
[out]	p_postave_mean	: posterior average of parameter mean
[out]	p_postave_var	: posterior average of parameter variance
[out]	p_postvar_mean	: posterior variance of parameter mean
[out]	f_postave_mean	: posterior average of function prediction mean
[out]	f_postave_var	: posterior average of function prediction variance
[out]	f_postvar_mean	: posterior variance of function prediction mean
[out]	paramPCcfs	: each column is a vector of parameter PC coefficients corresponding to an MCMC sample from pchain the last column is the MAP value of parameter PC coefficients

LogPosterior()

double LogPosterior	(	Array1D< double > &	m,
		void *	mypost_void )

Log-posterior function given a vector of parameters (chain state) and a void* set of auxiliary variables.