ParameterSpace class

[dreamer2368]

ParameterSpace class is factored out as a new object, supporting multi-dimensional parameter space.

Initialization

ParameterSpace is initiated with a dict-type variable config, which specifies parameters:

param_space = ParameterSpace(config)

Specifying parameters

With yaml format, a typical config (for the examples of burgers1d) looks like:

parameter_space:
  parameters:
    - name: a
      min: 0.7
      max: 0.9
      test_space_type: uniform
      sample_size: 21
      log_scale: false
    - name: w
      min: 0.9
      max: 1.1
      # another way of specifying test space.
      test_space_type: list
      list: [0.9, 0.91, 0.92, 0.93, 0.94,
             0.95, 0.96, 0.97, 0.98, 0.99,
             1.0, 1.01, 1.02, 1.03, 1.04,
             1.05, 1.06, 1.07, 1.08, 1.09, 1.1]
  test_space:
    type: grid

parameters category provides a list of parameters we'd want to perform parametric sampling:

• name specifies the name of the parameter, so that a Physics solver can read it.
• min and max value specify the range of the parameter
• test_space_type: if the test space if specified, then create 1d parameter space based on this type specification.
  • uniform: uniform grid space between min and max with sample_size points.
  • list: uses the specified list of parameter values.

test_space specifies whether or not to use FOM solutions on parameter space for test, and what type of test space we'd use. Currently it only supports grid type. Will have to support PR #3 .

grid-type space

By default, ParameterSpace creates the initial training samples and a test space as multi-dimensional grids. The dimension of parameter space is determined by the number of parameters specified in the input config.

passing parameters to Physics

ParameterSpace.getParameter function converts a numpy 1d array of parameter into a dict, so that Physics class can use parameters according to their names. Physics class can use default parameter values if not all parameter values are provided.

[chldkdtn]

Looks great!

[andersonw1]

Looks good! As you mentioned, in the future I think it would be helpful to add what I did in PR #3 where we provide the outline of the parameter space and put a grid over the convex hull.