Design of Experiment

When using Bayesian Optimization, HyperMapper always starts its optimization with a Design of Experiment (DoE) phase. In the DoE phase, HyperMapper selects a number of samples to evaluate in order to initialize its models, before starting the bayesian optimization loop. How HyperMapper chooses the points to explore during DoE depends on the type of DoE used. Currently, HyperMapper supports four different DoE methods: Random Sampling, Standard Latin Hypercube, K Latin Hypercube, and Grid Search. The DoE method used and the number of samples drawn during the DoE phase are controlled by the design_of_experiment field in the json scenario file:

"design_of_experiment": {
    "doe_type": "random sampling",
    "number_of_samples": 1000
  }

In HyperMapper's json schema, the DoE methods are called:

• random sampling
• standard latin hypercube
• k latin hypercube
• grid_search The number_of_samples field can be any positive integer. For more details, see the Json Parameters page.

Grid Search DoE

The grid search DoE has a different behavior compared to the other design of experiment methods. First, the grid search can only be used as a standalone method of exploring the input parameter space. If grid search is used as design of experiment, HyperMapper will automatically set the number of bayesian optimization iterations to 0. Second, grid search always explores all of the possible parameter combinations in the search space, ignoring the number_of_samples field in the design_of_experiment object. Note that this can lead to a very large number of function evaluations, depending on the size of the parameter space.

Note. For real parameters, grid search will discretize the parameter range into 10 equally spaced values. If a different discretization is preferred, we recommend encoding the real parameter as an ordinal or categorical parameter instead.