adapt_rank

Activate adaptive procedure for determining best rank representation

Specification

  • Alias: None

  • Arguments: None

  • JSON argument key: None

Description

The adaptive algorithm proceeds as follows:

  1. Start from rank start_rank and form an approximation

  2. Adapt the current approximation by searching for a solution with lower rank that achieves L2 accuracy within epsilon tolerance of the reference.

  3. If a lower rank solution is found with comparable accuracy, then stop. If not, increase the rank by an amount specified by kick_rank.

  4. Return to step 2 and continue until either max_rank is reached or a converged rank (rank less than current reference with comparable accuracy) is found.

Default Behavior

No cross validation for rank.

Examples

This example shows specification of a rank adaptation starting at rank 2, incrementing by 2, and limited at rank 10.

model,
 id_model = 'FT'
 global_surrogate function_train
   start_order = 5
   adapt_rank  start_rank = 2  kick_rank = 2  max_rank = 10
   solver_tolerance   = 1e-12
   rounding_tolerance = 1e-12
 dace_method_pointer = 'SAMPLING'
{
  "model": [
    {
      "global_surrogate": {
        "id_model": "FT",
        "function_train": {
          "start_order": 5,
          "adapt_rank": {
            "start_rank": 2,
            "kick_rank": 2,
            "max_rank": 10
          },
          "solver_tolerance": 1e-12,
          "rounding_tolerance": 1e-12
        },
        "dace_method_pointer": "SAMPLING"
      }
    }
  ]
}

Note that adapt_rank and adapt_order can either be combined or used separately.