MC Task

Overview

The base MCTask class at maestro/mc/mctask.py contains useful utility functions that will allow you to interface with the MÆSTRO algorithm with ease. Here, we describe the functions that need to be implemented in your class that inherits MCTask, a useful utility function to deal with NaNs/\(\infty\) in the MC output data, and the code documentation from MCTask class that describe the functions already defined in this class.

Function inherited from but not defined in the MCTask base class

The following functions need to be implemented in your class that inherits MCTask. If these functions are not defined, when they are called, it will result in an exception.

  • run_mc(self): This function is used to execute the MC task as a function call. See calling mc task as a function call tutorial for an example.

  • merge_statistics_and_get_max_sigma(self): This function is used to merge statistics (for multi-fidelity runs) and get the maximum standard deviation across all parameters from the MC run.

  • convert_mc_output_to_df(self, all_param_directory): Convert the output from the MC simulator for all parameters in the all_param_directory to a pandas data frame output. An example dataframe output is shown below. In this example, there are three parameters and two terms of the objective function. Terms that end with .P are the parameters and those ending with .V are the values associated with either the MC, i.e., MC sample values or the DMC, i.e., MC standard deviation values. You can add more rows, i.e, more sets of parameter and values for additional terms in the objective function or more columns, i.e., more components of the each term of the objective that come from the MC simulator.:

    >df
                MC                      DMC
Term1.P        [[1,2],[3,4],[6,3]]     [[1,2],[3,4],[6,3]]
Term1.V        [19, 18, 17]            [99, 98, 97]
Term2.P        [[1,2],[3,4],[6,3]]     [[1,2],[3,4],[6,3]]
Term2.V        [29, 28, 27]            [89, 88, 87]

Dealing with NaNs/\(\infty\) in MC output

To deal with NaNs/\(\infty\) in MC output, we provide a useful utility function:

def check_and_resolve_nan_inf(self,data, all_param_directory,term_names,group_names=None)

The strategy used to deal with NaNs/\(\infty\) in MC output is as follows:

  • If group_names is not None i.e., groups are specified, then try to interpolate NaNs/\(\infty\) for terms within each group for each parameter

  • If group_names is None i.e., groups are not specified or if interpolation does not completely resolve the issue, then:

    • If the NaNs/\(\infty\) was encountered in the start parameter, then raise an exception

    • If the NaNs/\(\infty\) was encountered in the iterate (subproblem argmin), then use the surrogate models from the current iteration to fill in the missing values

    • If the NaNs/\(\infty\) was encountered in the adaptive sample, then drop the parameter from the corresponding term

More information can be found in the MC Task Code documentation.

MC Task code documentation

See MC Task API description in the code documentation section for a complete documentation of all implemented and abstract (not implemented) functions within MCTask.