DAGMANonlinear<T>

Models & Types · AiDotNet.CausalDiscovery.ContinuousOptimization

DAGMA Nonlinear — DAG learning via M-matrices and log-determinant with MLP structural equations.

Inheritance: CausalDiscoveryBase<T> → ContinuousOptimizationBase<T> → DAGMANonlinear<T>

For Beginners

This is like DAGMA Linear but can discover non-linear causal relationships using small neural networks. While DAGMA Linear assumes Y = a*X + b, this version can learn curved relationships like Y = sin(X) or Y = X^2.

How It Works

Extends DAGMA to nonlinear settings by parameterizing each structural equation with a small MLP. Uses the same log-determinant acyclicity constraint as DAGMA Linear but applied to the adjacency matrix extracted from MLP input-layer weights.

Constraint: h(theta, s) = -log det(sI - A∘A) + d*log(s), where A[i,j] = ||W1_j[:,i]||_2

Reference: Bello et al. (2022), “DAGMA: Learning DAGs via M-matrices and a Log-Determinant Acyclicity Characterization”, NeurIPS.

Constructors

Constructor Summary
DAGMANonlinear(CausalDiscoveryOptions options)

Properties

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

Methods

Method Summary
ValueTuple<int, double, double> GetLastRunInfo() Gets the iteration count and convergence info from the last run.