kl_recursion

Model graph enumeration that follows the ACV-KL partitioning scheme

Specification

  • Alias: None

  • Arguments: None

  • JSON argument key: None

  • Default: true

Description

The kl_recursion approach (known as ACV-KL in [GGEJ20]) enumerates a set of free parameters, K and L, that partition a sequence of control variate targets within the directed acyclic graph (DAG) of control variate pairings. Within a set of approximation models of size M, the K highest-fidelity approximations target the root node (the truth model) whereas the other M-K lowest fidelity approximations all target node L. The DAGs associated with different values for K and L are enumerated and the one with the best performance (lowest estimator variance for a prescribed budget or lowest cost for a prescribed accuracy) is selected for final post-processing.

As described in [BLWL22], the kl_recursion approach defines a subset of ordered DAGs that are contained within the unordered subset of DAGs defined by partial_recursion depth_limit = 2. As such, it explores a reduced number of alternatives which may be appropriate for larger model ensembles.

Examples

method
  model_pointer = 'ENSEMBLE'
  approximate_control_variate acv_mf
    pilot_samples = 25 seed = 8674132
    search_model_graphs kl_recursion
    max_function_evaluations = 500
{
  "method": [
    {
      "approximate_control_variate": {
        "model_pointer": "ENSEMBLE",
        "pilot_samples": [25],
        "seed_sequence": [8674132],
        "max_function_evaluations": 500
      }
    }
  ]
}

Theory

Refer to [GGEJ20] for additional details.