DNN

Purpose

DNN performs iterative surrogate-assisted sampling: initial true evaluations, neural-network training, recommendation, then repeated refinement.

Full Sampling Section Keys

• Sampling.Method (required): must be DNN.
• Sampling.Variables (required, array):
• name, description, distribution.type, distribution.parameters
• Runtime-safe parameter sets:
  • Flat: min, max
  • Log: min, max
  • Normal: mean, stddev
  • Log-Normal: mean, stddev
  • Logit: location, scale
• Sampling.LogLikelihood (required, array): {name, expression}
• Sampling.selection (optional, string): candidate filter expression.
• Sampling.Bounds (required, object):
• Niters (required, integer): outer DNN iterations
• Hidden_layers (required, array of integers)
• Batch_size (required, integer)
• Ninit (required, integer)
• Nepoch (required, integer)
• Learning_rate (required, number)
• Outputs (required, array of strings)
• Prop_new (optional, number, default runtime 0.1)

Full Skeleton

Sampling:
  Method: "DNN"
  Variables:
    - name: p1
      description: parameter 1
      distribution:
        type: Flat
        parameters:
          min: 0.0
          max: 1.0
  LogLikelihood:
    - name: L_total
      expression: "-0.5*((obs-100.0)/10.0)^2"
  selection: "p1 > 0"
  Bounds:
    Niters: 20
    Hidden_layers: [128, 128, 64]
    Batch_size: 128
    Ninit: 2000
    Nepoch: 200
    Learning_rate: 0.001
    Outputs: [obs]
    Prop_new: 0.15

Example

Sampling:
  Method: "DNN"
  Variables:
    - name: xx
      description: x
      distribution:
        type: Flat
        parameters:
          min: 0.0
          max: 31.4159
    - name: yy
      description: y
      distribution:
        type: Flat
        parameters:
          min: 0.0
          max: 31.4159
  LogLikelihood:
    - name: L_z
      expression: "-0.5*((z-100.0)/10.0)^2"
  selection: "xx > 0 and yy > 0"
  Bounds:
    Niters: 20
    Hidden_layers: [128, 128, 64]
    Batch_size: 128
    Ninit: 2000
    Nepoch: 200
    Learning_rate: 0.001
    Outputs: [z]
    Prop_new: 0.15