Batched Evaluations

Dakota provides for execution of evaluations in batches. Batch mode is intended to allow a user to assume greater control over where and when to run individual evaluations. It is activated using the batch keyword.

In batch mode, Dakota writes the parameters for multiple (a batch of) evaluations to a single batch parameters file and then invokes the analysis driver once for the entire batch. The pathname of the combined parameters file (and of the results file) are communicated to the driver as command line arguments. After the driver exits, Dakota expects to find results for the entire batch in a single combined results file.

The analysis driver is responsible for parsing the parameters file and performing an evaluation for each set of parameters it contains, and for returning results for all the evaluations to Dakota. The user is free to set up the driver to perform the evaluations in the batch in a way that is convenient.

By default, all currently available evaluations are added to a single batch. For example, in a sampling study that has a 1000 samples, by default all 1000 evaluations would be added to a single batch. The batch size may be limited using the size subkeyword. Setting size = 100 would result in 10 equal-size batches being run one after another in a 1000-sample study.

File Formats

For standard and APREPRO-format parameters files, the combined parameters file for a batch is simply the concatenation of the paramters information for all the evaluations in the batch, in standard or APREPRO format.

The following example parameters file contains parameter sets for two evaluations.

                     1 variables
-4.912558193411678e-01 x1
                     1 functions
                     1 ASV_1:response_fn_1
                     1 derivative_variables
                     1 DVV_1:x1
                     0 analysis_components
                   1:1 eval_id
                     1 variables
-2.400695372000337e-01 x1
                     1 functions
                     1 ASV_1:response_fn_1
                     1 derivative_variables
                     1 DVV_1:x1
                     0 analysis_components
                   1:2 eval_id    interface,

Note that the eval_id contains two pieces of information separated by a colon. The second is the evaluation number, and the first is the batch number. The batch number is an incrementing integer that uniquely identifies the batch.

A JSON format batch parameters file contains an array at the top level. The elements of the array are objects that contain parameters information for each evaluation in the batch.

The combined results file format is likewise a concatenation of the results for all the evaluations in the batch. However, a line beginning with the “#” character must separate the results for each evaluation.

The order of the evaluations in the results file must match the order in the parameters file.

The following is an example batch results file corresponding to the batch parameters file above. The initial # on the first line is optional, and a final # (not shown here) is allowed.

#
                     4.945481774823024e+00 f
#
                     2.364744129789246e+00 f

Work Directories, Tagging, and Other Features

Each batch is executed in a work directory when this feature is enabled. The batch number is used to tag files and directories if tagging is requested (or Dakota automatically applies a tag to safely save a file or directory). As explained in the previous section, the batch number is an incrementing integer beginning with 1 that uniquely identifies a batch.

Batch mode restricts the use of several other Dakota features:

• No input_filter or output_filter is allowed.

• Only one analysis_drivers keyword is permitted.

• failure_capture modes are limited to abort and recover.

• Asynchronous evaluation is disallowed (only one batch at a time may be executed).