IterativeMCMCAlgorithm<T>

Models & Types · AiDotNet.CausalDiscovery.Bayesian

Iterative MCMC — iteratively refined MCMC for Bayesian network structure learning.

Inheritance: CausalDiscoveryBase<T> → BayesianCausalBase<T> → IterativeMCMCAlgorithm<T>

For Beginners

Standard MCMC can get “stuck” in local optima. Iterative MCMC uses clever tricks to escape these traps and explore more of the solution space, leading to better posterior estimates.

How It Works

Iterative MCMC improves MCMC mixing by alternating between different proposal mechanisms: edge additions, deletions, and reversals. It uses multiple restarts with the best DAG from each restart used to seed the next, progressively refining the search.

Algorithm:

  • Initialize DAG from empty graph or previous restart’s best
  • Propose: randomly add, delete, or reverse an edge
  • Check acyclicity of proposed DAG
  • Compute BIC score difference
  • Accept/reject via Metropolis-Hastings
  • After burn-in, accumulate edge posterior probabilities
  • Restart with best DAG found so far as seed

Reference: Kuipers et al. (2017), “Efficient Structure Learning and Sampling of Bayesian Networks”, arXiv.

Constructors

Constructor Summary
IterativeMCMCAlgorithm(CausalDiscoveryOptions options)

Properties

Property Summary
bool SupportsNonlinear { get; }
string Name { get; }