void MBC_MCMC_loop(ERGMM_MCMC_Model *model, ERGMM_MCMC_Priors *prior,
ERGMM_MCMC_MCMCState *cur, ERGMM_MCMC_MCMCSettings *setting, ERGMM_MCMC_ROutput *outlists)
{
unsigned int pos=0;
unsigned int iter, total_iters = setting->sample_size*setting->interval;
// Rprintf("Started MCMC loop.\n");
/* Note that indexing here starts with 1.
It can be thought of as follows:
At the end of the updates, we have made iter complete MCMC updates. */
for(iter=1;iter<=total_iters;iter++){
R_CheckUserInterrupt(); // So that CTRL-C can interrupt the run.
ERGMM_MCMC_CV_up(model,prior,cur);
ERGMM_MCMC_logp_Z(model, cur->state);
// If we have a new MLE (actually, the highest likelihood encountered to date), store it.
if( cur->state->lpZ > outlists->lpZ[0] ) MBC_MCMC_store_iteration(0,model,cur->state,setting,outlists);
if( cur->state->lpZ + cur->state->lpLV > outlists->lpZ[1] + outlists->lpLV[1] ) MBC_MCMC_store_iteration(1,model,cur->state,setting,outlists);
/* every interval save the results */
if((iter % setting->interval) == 0){
pos = (iter/setting->interval-1)+MBC_OUTLISTS_RESERVE;
// Store current iteration.
MBC_MCMC_store_iteration(pos, model, cur->state, setting, outlists);
}
} // end main MCMC loop
return;
}
/* Initializes the MCMC sampler and allocates memory. */
void ERGMM_MCMC_init(unsigned int sample_size, unsigned int interval,
unsigned int n,
unsigned int p, unsigned int d, unsigned int G,
unsigned int dir, int **iY, double **dY,
unsigned int family,
int *iconsts, double *dconsts,
unsigned int latent_eff,
double ***X,
double *llk_mcmc, double *lpZ_mcmc, double *lpcoef_mcmc, double *lpRE_mcmc, double *lpLV_mcmc, double *lpREV_mcmc, double *lpdispersion_mcmc,
double **Z_start,
double *Z_pK_start, double **Z_mean_start, double *Z_var_start, unsigned int *Z_K_start,
double Z_var_prior, double Z_mean_prior_var, double Z_pK_prior,
double Z_var_prior_df,
double *Z_mcmc, double *Z_rate_move, int *K_mcmc,
double *Z_pK_mcmc,
double *Z_mean_mcmc, double *Z_var_mcmc,
double *coef_start,
double *coef_mcmc, double *coef_rate,
double *coef_prior_mean, double *coef_var,
double *sender_start, double *receiver_start,
double sender_var_start, double receiver_var_start,
double sender_var_prior, double sender_var_prior_df,
double receiver_var_prior, double receiver_var_prior_df,
double *sender_mcmc, double *receiver_mcmc,
double *sender_var_mcmc, double *receiver_var_mcmc,
double dispersion_start, double dispersion_prior,
double dispersion_prior_df, double *dispersion_mcmc,
unsigned int sociality,
unsigned int **observed_ties,
double Z_delta,
double RE_delta,
double **group_deltas,
unsigned int group_prop_size,
double **coef_eff_sender,
unsigned int coef_eff_sender_size,
double **coef_eff_receiver,
unsigned int coef_eff_receiver_size,
unsigned int accept_all)
{
unsigned int i;
// Packing constants into structs.
ERGMM_MCMC_Model model = {dir,
iY, // iY
dY, // dY
X, // X
(unsigned int **)observed_ties,
ERGMM_MCMC_lp_edge[family],
ERGMM_MCMC_E_edge[family],
0,
iconsts,
dconsts,
n, // verts
d, // latent
p, // coef
G, // clusters
sociality,
dispersion_mcmc!=NULL,
d ? ERGMM_MCMC_latent_eff[latent_eff] : NULL
};
ERGMM_MCMC_set_lp_Yconst[family](&model);
ERGMM_MCMC_MCMCSettings setting = {Z_delta,
RE_delta,
group_deltas,
coef_eff_sender,
coef_eff_receiver,
group_prop_size,
coef_eff_sender_size,
coef_eff_receiver_size,
sample_size,interval,
accept_all
};
ERGMM_MCMC_Priors prior = {Z_mean_prior_var, // Z_mean_var
Z_var_prior, // Z_var
Z_var_prior_df, // a.k.a. Z_var_df (I hope)
coef_prior_mean,
coef_var,
Z_pK_prior,
sender_var_prior,
sender_var_prior_df,
receiver_var_prior,
receiver_var_prior_df,
dispersion_prior,
dispersion_prior_df};
ERGMM_MCMC_Par state = {Z_start, // Z
coef_start, // coef
Z_mean_start, // Z_mean
Z_var_start, // Z_var
Z_pK_start, // Z_pK
sender_start,
sender_var_start,
model.sociality?sender_start:receiver_start,
receiver_var_start,
dispersion_start,
Z_K_start, // Z_K
0, // llk
dmatrix(model.verts,model.verts), // lpedge
0, // lpZ
0, // lpLV
0, // lpcoef
0, // lpRE
0, // lpREV
0, // lpdispersion
};
ERGMM_MCMC_Par prop = {model.latent ? dmatrix(model.verts,model.latent):NULL, // Z
model.coef ? dvector(model.coef):NULL, // coef
model.clusters ? dmatrix(model.clusters,model.latent):NULL, // Z_mean
model.latent ? dvector(model.clusters?model.clusters:1):NULL, // Z_var
model.clusters ? dvector(model.clusters):NULL, // Z_pK
sender_start ? dvector(model.verts):NULL, // sender
0, // sender_var
receiver_start && !model.sociality ? dvector(model.verts):NULL, // receiver
0, // receiver_var
0, // dispersion
state.Z_K, // prop.Z_K === state.Z_K
0, // llk
dmatrix(model.verts,model.verts), // lpedge
0, // lpZ
0, // lpLV
0, // lpcoef
0, // lpRE
0, // lpREV
0 // lpdispersion
};
if(model.sociality) prop.receiver=prop.sender;
ERGMM_MCMC_MCMCState start = {&state,
&prop,
model.clusters ? dmatrix(model.clusters,model.latent) : NULL, // Z_bar
setting.group_prop_size ? dvector(setting.group_prop_size) : NULL, // deltas
model.clusters ? dvector(model.clusters): NULL, // pK
model.clusters ? (unsigned int *) ivector(model.clusters) : NULL, // n
PROP_NONE, // prop_Z
PROP_NONE, // prop_RE
PROP_NONE, // prop_coef
PROP_NONE, // prop_LV
PROP_NONE, // prop_REV
PROP_NONE, // prop_dispersion
FALSE, // after_Gibbs
(model.latent || sender_start || receiver_start) ? (unsigned int *) ivector(model.verts) : NULL // update_order
};
ERGMM_MCMC_ROutput outlists = {llk_mcmc, lpZ_mcmc, lpcoef_mcmc, lpRE_mcmc, lpLV_mcmc, lpREV_mcmc, lpdispersion_mcmc,
Z_mcmc, Z_rate_move,
coef_mcmc,coef_rate,
Z_mean_mcmc,Z_var_mcmc,Z_pK_mcmc,
sender_mcmc,sender_var_mcmc,
receiver_mcmc,receiver_var_mcmc,
dispersion_mcmc,
K_mcmc};
if(model.clusters>0)
for(i=0;i<model.verts;i++)
start.n[state.Z_K[i] - 1]++;
// Initialize the log-probabilities.
state.llk = ERGMM_MCMC_lp_Y(&model, &state, TRUE);
if(model.latent) ERGMM_MCMC_logp_Z(&model, &state);
if(state.sender || state.receiver) ERGMM_MCMC_logp_RE(&model, &state);
if(state.coef) ERGMM_MCMC_logp_coef(&model, &state, &prior);
if(model.latent) ERGMM_MCMC_logp_LV(&model, &state, &prior);
if(state.sender || state.receiver) ERGMM_MCMC_logp_REV(&model, &state, &prior);
if(model.dispersion) ERGMM_MCMC_logp_dispersion(&model, &state, &prior);
copy_MCMC_Par(&model,&state,&prop);
ERGMM_MCMC_store_iteration(0,&model,&state,&setting,&outlists);
ERGMM_MCMC_store_iteration(1,&model,&state,&setting,&outlists);
ERGMM_MCMC_loop(&model, &prior, &start, &setting, &outlists);
// R's garbage collector takes care of freeing memory.
}
/* Initializes the MCMC sampler and allocates memory. */
void MBC_MCMC_init(unsigned int sample_size,
unsigned int interval,
unsigned int n,
unsigned int d,
unsigned int G,
double *lpZ_mcmc,
double *lpLV_mcmc,
double **Z,
double *Z_pK,
double **Z_mean_start,
double *Z_var,
unsigned int *Z_K,
double Z_var_prior,
double Z_mean_prior_var,
double Z_K_prior,
double Z_var_df,
int *K_mcmc,
double *Z_pK_mcmc,
double *Z_mean_mcmc,
double *Z_var_mcmc){
unsigned int i;
// Packing constants into structs.
ERGMM_MCMC_Model model = {0,
NULL, // iY
NULL, // dY
NULL, // X
NULL,
NULL,
NULL,
0,
NULL,
NULL,
n, // verts
d, // latent
0, // coef
G // clusters
};
ERGMM_MCMC_MCMCSettings setting = {0,0,NULL,NULL,NULL,0,0,0, // deltas
sample_size,interval,
FALSE // accept_all
};
ERGMM_MCMC_Priors prior = {Z_mean_prior_var, // Z_mean_var
Z_var_prior, // Z_var
Z_var_df, // a.k.a. Z_var_df (I hope)
NULL,
NULL,
Z_K_prior,
0,0,0,0};
ERGMM_MCMC_Par state = {Z, // Z
NULL, // coef
Z_mean_start, // Z_mean
Z_var, // Z_var
Z_pK, // Z_pK
NULL,
0,
NULL,
0,
0, // dispersion
Z_K, // Z_K
0, // llk
NULL, // lpedge
0, // lpZ
0, // lpLV
0, // lpcoef
0, // lpRE
0, // lpREV
0 // lpdispersion
};
ERGMM_MCMC_MCMCState start = {&state,
NULL,
model.clusters ? dmatrix(model.clusters,model.latent) : NULL, // Z_bar
NULL, // tr_by
model.clusters ? dvector(model.clusters): NULL, // pK
model.clusters ? (unsigned int *) ivector(model.clusters) : NULL, // n
PROP_NONE, // prop_Z
PROP_NONE, // prop_RE
PROP_NONE, // prop_coef
PROP_NONE, // prop_LV
PROP_NONE, // prop_REV
PROP_NONE, // prop_dispersion
FALSE, // after_Gibbs
NULL // update_order
};
ERGMM_MCMC_ROutput outlists = {NULL, // llk
lpZ_mcmc,
NULL, // lpcoef
NULL, // lpRE
lpLV_mcmc,
NULL, //lpREV,
NULL, //lpdispersion
NULL, // Z
NULL, // Z_rate_move
NULL, // coef
NULL, // coef_rate
Z_mean_mcmc,Z_var_mcmc,Z_pK_mcmc,
NULL,NULL,
NULL,NULL,
NULL, // dispersion_mcmc
K_mcmc};
if(model.clusters>0)
for(i=0;i<model.verts;i++)
start.n[state.Z_K[i] - 1]++;
// Initialize the log-probabilities.
ERGMM_MCMC_logp_Z(&model, &state);
MBC_MCMC_store_iteration(0,&model,&state,&setting,&outlists);
MBC_MCMC_store_iteration(1,&model,&state,&setting,&outlists);
MBC_MCMC_loop(&model, &prior, &start, &setting, &outlists);
}
