NoCurlAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.ContinuousOptimization
NoCurl — DAG learning via curl-free constraints on the graph structure.
Inheritance: CausalDiscoveryBase<T> → ContinuousOptimizationBase<T> → NoCurlAlgorithm<T>
For Beginners
NoCurl avoids the expensive matrix exponential used by NOTEARS. Instead, it ensures acyclicity by finding a variable ordering (X before Y before Z) and only allowing edges that follow this ordering. This is faster and simpler while still finding good causal structures.
How It Works
NoCurl parameterizes the DAG using a variable ordering and restricted edge weights. Given an ordering pi, the adjacency matrix W is constrained to have edges only from earlier to later variables in the ordering — guaranteeing acyclicity by construction. The algorithm alternates between optimizing the ordering (via greedy swaps) and optimizing edge weights (via OLS regression restricted to the ordering).
Algorithm:
- Initialize with an ordering based on marginal variance
- For the current ordering, compute optimal edge weights via OLS (restricted to the ordering)
- Apply L1 soft-thresholding for sparsity
- Try adjacent swaps to improve the L2 loss
- Accept the best swap and repeat
- Threshold small weights in the final W
Reference: Yu et al. (2021), “DAGs with No Curl: An Efficient DAG Structure Learning Approach”, ICML.
Constructors
| Constructor | Summary |
|---|---|
NoCurlAlgorithm(CausalDiscoveryOptions options) |
Initializes NoCurl with optional configuration. |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |