Version 6.19 (2023/11/15)

Highlight: New sampling-based method for Sobol’ main effects

Based on [LM16], Dakota can now obtain estimates of first order Sobol’ indices (main effects) from sampling studies. Previous versions of Dakota could estimate main and total effects using a “pick and freeze” sampling strategy [STCR04], which typically required a very large number of samples (hundreds or thousands per variable) that had to be carefully structured. While the new method produces only main effects, the requirement on sample design has been lifted, and typically far fewer samples are needed for convergence.

Enabling / Accessing:

The variance_based_decomp keyword now has suboptions. The vbd_sampling_method pick_and_freeze option is the default, and causes Dakota to use the method that has long been available to compute main and total effects. The vbd_sampling_method binned option causes the new method to be used.

Documentation:

Keyword reference for the binned VBD method.

Highlight: Low-discrepancy (quasi-Monte Carlo) sampling

Two new strategies for choosing low-discrepancy points in sampling studies are available in this release. These include lattice rules and digital nets. The well-known Sobol sequence is an example of a digital net. Just as in Latin hypercube sampling, these strategies choose points that cover the parameter space more uniformly than ordinary Monte Carlo, leading to faster convergence of UQ results.

Enabling / Accessing:

In a sampling study, choose sample_type low_discrepancy.

Documentation:

Highlight: Model selection in multifidelity sampling methods

MFMC, ACV, and generalized ACV now support selection of the most performant subset of model approximations, as determined through enumeration by the estimator accuracy versus equivalent cost trade-off.

Enabling / Accessing:

In a multifidelity_sampling or approximate_control_variate study, choose search_model_graphs model_selection.

Documentation:

Improvements by Category

Interfaces, Input/Output

UQ Methods

  • Improved support for Bayesian calibration methods provided by the MIT Uncertainty Quantification Library (MUQ); although MUQ is not enabled in our pre-built downloads, it is now buildable within Dakota by a wider variety of toolchains.

  • Low discrepancy sampling strategies (see highlight)

MLMF Sampling

  • Support for selection of the most performant approximation subset, as highlighted above.

  • New options for more robust numerical solution of optimal sample allocations, including global_local and competed_local solution strategies. The former is now the default for numerical solutions in MFMC, ACV, and generalized ACV.

Sensitivity Analysis

  • Binned method for sampling-based variance-based decomposition (see highlight)

  • New examples explaining use of correlation coefficients for global sensitivity analysis.

Miscellaneous Enhancements and Bugfixes

  • Enh: Documentation of Dakota’s regression test system expanded.

  • Bug fix: The @python_interface decorator in the dakota.interfacing module now propertly interprets the dvv list provided by Dakota’s direct python interface.

  • Bug fix: RPATH handling on Linux-based platforms improved.

  • Enh: pyprepro gained a new function, json_dumps(), which returns all variables (and their values) formatted as a JSON string. See the pyprepro manual for more information.

Compatibility

  • Support for building Dakota with C++17 has been greatly expanded and is expected to work for GCC, Intel, and Clang compilers. Support for Microsoft Visual Studio in progress.

Known Issues

  • Plotting in the Dakota GUI on RHEL8 may fail with the error “Unhandled event loop exception. No more handles because there is no underlying browser available.” There currently is no known workaround or resolution.