NUTS

Purpose

NUTS is an adaptive-depth HMC variant (No-U-Turn Sampler).

Full Sampling Section Keys

• Sampling.Method (required): must be NUTS.
• Sampling.Variables (required):
• 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 of {name, expression}
• Sampling.selection (optional, string)
• Sampling.Bounds:
• base keys: num_chains, num_iters, proposal_scale
• NUTS keys:
  • nuts_step_size (optional, number, default 0.05)
  • nuts_max_depth (optional, integer, default 6)
  • step_size (optional, number, alias)
  • max_depth (optional, integer, alias)

Full Skeleton

Sampling:
  Method: "NUTS"
  Variables:
    - name: x
      description: variable x
      distribution:
        type: Flat
        parameters:
          min: -5
          max: 5
  LogLikelihood:
    - name: L_x
      expression: "-0.5*(x/1.0)^2"
  Bounds:
    num_chains: 6
    num_iters: 9000
    proposal_scale: [0.05, 0.05, 0.05, 0.05, 0.05, 0.05]
    nuts_step_size: 0.03
    nuts_max_depth: 7
    step_size: 0.03
    max_depth: 7

Example

Sampling:
  Method: "NUTS"
  Variables:
    - name: x
      description: variable x
      distribution:
        type: Flat
        parameters:
          min: -5
          max: 5
  LogLikelihood:
    - name: L_x
      expression: "-0.5*(x/1.0)^2"
  Bounds:
    num_chains: 6
    num_iters: 9000
    proposal_scale: [0.05, 0.05, 0.05, 0.05, 0.05, 0.05]
    nuts_step_size: 0.03
    nuts_max_depth: 7