ergmx.Control

class ergmx.Control(samplesize=1024, interval=1024, burnin=16384, n_chains=<factory>, triadic_weight=None, effective_size=64, max_interval=1048576, max_iter=60, steplength_margin=0.05, last_boost=4, density_guard=20.085536923187668, density_guard_min=10000, stall_iterations=10, cd_steps=8, cd_samplesize=1024, cd_max_iter=60, cd_conv_min_pval=0.5, bridges=32, bridge_chains=1, bridge_samplesize=256)

Tuning parameters of the MCMC and of the Monte Carlo MLE.

The defaults follow ergm: 1024 samples, 1024 proposals apart, after 16 times as many burn-in proposals. When the samples are too autocorrelated to reach effective_size, the interval (and the burn-in) grows.

samplesize = 1024

Networks sampled per iteration, split across the chains.

interval = 1024

MCMC proposals between two sampled networks, at the start.

burnin = 16384

MCMC proposals discarded at the start of every chain, at the start.

n_chains

Parallel Markov chains (one thread each).

triadic_weight = None

Share of triadic MCMC proposals (they close or open a triangle). None uses 0.5 for models with triangle or shared partner terms, 0 otherwise.

effective_size = 64

Minimum effective sample size per iteration; None keeps the interval fixed.

max_interval = 1048576

Largest interval the adaptation may reach.

max_iter = 60

Maximum number of Monte Carlo MLE iterations.

steplength_margin = 0.05

Margin by which the target must lie inside the sample’s convex hull.

last_boost = 4

The final iteration samples this many times more networks.

density_guard = 20.085536923187668

Stop if a simulated network has more than this many times the observed edges (and more than density_guard_min), as ergm does.

density_guard_min = 10000

Edges a simulated network may always have, whatever the density guard.

stall_iterations = 10

Stop after this many consecutive iterations with a step length below 0.1; None never stops early.

cd_steps = 8

MCMC proposals from the observed network per sample.

Type:

Contrastive divergence

cd_samplesize = 1024

samples per iteration.

Type:

Contrastive divergence

cd_max_iter = 60

maximum number of iterations.

Type:

Contrastive divergence

cd_conv_min_pval = 0.5

Contrastive divergence stops when Hotelling’s test of the samples against the observed statistics has a larger p-value, as in ergm.

bridges = 32

intervals along the path, with samples at both ends of each.

Type:

Log-likelihood by path sampling

bridge_chains = 1

chains per point of the path.

Type:

Log-likelihood by path sampling

bridge_samplesize = 256

samples per chain, spaced by the interval the Monte Carlo MLE ended with.

Type:

Log-likelihood by path sampling