Version 6.3 (2015/11/16)

Highlights

• Numerous algorithmic and usability enhancements to Bayesian calibration capabilities.

• Gaussian process models now exportable in human-readable format.

• Incremental LHS now works for all discrete variable types including histograms and discrete sets. Support for string variables was also extended to the JEGA, COLIN, and NOMAD optimizers.

• Updated training materials made available in the Community->Training section of the Dakota website.

• Improvements in computation and consistency of reporting of PDFs and CDFs across UQ methods.

• New interface keyword ‘labeled’ enables more rigorous results file validation and error reporting.

Uncertainty Quantification (UQ)

• UQ Statistics:

  • Improved computation of PDF values and empirical histogram generation in sampling methods (also PCE/SC and L/G reliability, IS, POF).

  • Generalize nonD stats compute/print to accommodate response and variable ensembles

  • Local reliability: improved computation of importance factors.

Optimization and Calibration

• Increased support for discrete variables

  • All types of discrete variables (integer, real, string) now mapped through to and optimized over by all evolutionary algorithms (soga, moga, coliny_ea) and asynchronous pattern search (asynch_pattern_search)

  • Experimental branch-and-bound capability added to optimize over mixed continuous-discrete variables when discrete variables can be relaxed

• Consistent support for multiple experiments across least squares calibration, optimizer, and Bayesian methods. Significant refactor of data and scaling transformations to allow options to work in combination with each other.

• Bayesian calibration algorithms and architecture:

  • Surrogate adaptive preconditioned MCMC using Hessian (or gradient) of simulation or emulator

    • New Hessian preconditioning based on eigenvalue truncation

  • MAP pre-solve option using deterministic optimizer, including for error hyper-parameters

  • QUESO now works in when invoking Dakota in serial or parallel execution

  • Significant improvements to file I/O and final results reporting, including ability to perform forward UQ based on a posterior chain.

  • Consistent support for PCE regression with least interpolation across Bayesian methods

  • Observation errors

    • When providing experimental data, users can also specify the covariance of the associated observation error process

    • Users may calibrate one or more hyper-parameters multiplying the covariance of the error, where error model hyper-parameters have inverse gamma prior

  • Update to QUESO v0.54

• Field data use in calibration:

  • Changed the way coordinates are read for simulation field responses to be similar to the way they are specified for experimental field data

  • Now, specify “read_field_coordinates” to read the coordinate file for simulation data, with the file name in the format of “response_descriptor.coords.”

  • Improved error handling when reading covariance data for the errors in experimental observations

Methods (general)

• Implemented new Bootstrap process and the Luo/Li 2015 “ladle” diagnostic to help automatically decide how many principal directions to include in the reduced space model. Added a new simple verification problem with a known randomly generated subspace of user-controlled size.

• Incremental LHS now works for all discrete variable types including histograms and discrete sets. Leverages new RandomVariable capabilities in Pecos.

• Voronoi Decomposition:

  • Improved domain decomposition capabilities of global surrogate models. Models based on polynomial regression can now use an integer basis order (0, 1, 2, 3 …) as well as classically defined keywords (linear, quadratic, cubic).

  • Domain-decomposed global surrogate models can now take advantage of derivatives’ information (gradients/Hessians), if available.

Framework

Input/Output

• Global I/O generalizations to distinguish point sets: import_build_points_file (training points upon which to build an approximation), export_approx_points_file (prediction points from an approximation or surrogate model), import_challenge_points_file (points at which to evaluate the surrogate model)

  • Clarify and standardize existing I/O for NonDBayesCalibration

    • Remove special option for surrogate point export

    • Add support for build point import to PCE and SC, consistent with GP

    • Ensure consistency in PCE spec among QUESO, DREAM, WASABI

  • Exports acceptance chain in user space

  • Additional chain statistics added

  • QUESO output directory renamed to QuesoDiagnostics and better described

• Forward propagation of arbitrary sample sets

  • File import_approx_points_file added for PCE/SC allowing evaluation of user-specified point sets on the surrogate model

  • Added format support and test coverage in PCE/SC

Variables

Add new RandomVariable components in Pecos to manage probability distributions. Use the new code to improve incremental LHS and to use log prior values and Hessians in QUESO Bayesian calibration.

Build System / TPLs

• LHS: fix compile error on array sizes, fix bug in RNG precision

• Surfpack Boost serialization bug with newer compilers.

• Suppress Boost signals deprecation warning in Acro

User Experience

• Top-level method controls (such as max_iterations, convergence_tolerance) are now properly associated in dakota.xml with the methods that support them, reducing user input errors/surprises. Reference manual now generated directly from dakota.xml, including flowing default values and more helpful alternation group labels.

• Surrogate Model Export

  • Can now export surfpack gaussian process models in “algebraic” format (augmenting existing ability to export polynomials, neural networks, and radial basis functions)

  • Full support for output options with filename to text, binary, algebraic file/console

• Update training materials (slides, examples, exercises) to reflect a more analyst-centric view and to address difficulties encountered by new users. Added training materials to the Dakota website.

• Improved validation of results files

  • Increased clarity of results file-related error messages

  • New interface keyword ‘labeled’ enables stricter checking and more verbose error reporting for results files; Requires that function values be correctly labeled with their descriptors

  • Variable and response descriptors no longer permitted to contain whitespace or to resemble floating point numbers

• ENH: dprepro allows c-style format specifiers on a per-tag basis

• Removed duplicate user manual tests and automatically generate user manual examples in the User’s Manual.

Miscellaneous Enhancements

Architecture

• ENH: String variables now available in direct interface, together with a textbook string variables tester

• ENH: Work directories are now uniquely tagged to work with concurrent methods in MPI mode

• ENH: Acro and DDACE cmake config files moved to new directory to better integrate with CASL VERA

Examples / Tests

• Improvements to built-in test drivers

  • Bayes linear tester for testing correctness of inferred posterior parameter distributions.

  • Flexibility in cantilever testers, added two higher dimensional rosenbrock: generalized (sums of coupled 2-D Rosenbrock functions in the objective) and extended (sums of uncoupled 2-D Rosenbrock functions).

  • Damped harmonic oscillator test driver: Returns an analytical time-dependent solution of a damped harmonic oscillator. The problem takes as input 1-6 random variables and returns the solution at a pre-specified number of equidistant time points.

  • text_book function extended to accept an arbitrary number of discrete string variables.

  • 1D (spatial) diffusion equation with random coefficients. The problem takes as input d>1 random variables which are coefficients of a KLE like diffusivity field and returns the spatial solution at a pre-specified number of equidistant spatial locations.

• Improvements to cross-platform test performance

  • Changed 30 tests from fork/system to direct interface to reduce testing time and cross-platform differences

  • Removed 36 duplicate user manual tests

  • 58 cross-platform improvements and 26 small regressions in eval counts

Miscellaneous Bug Fixes

• BUG: COBYLA optimization was ignoring max fn evals

• BUG: COLINY Beta supports integer domains

• BUG: patches Teuchos SerialSymDenseMatrix copy constructor

• BUG: communicator init/set/free would fail when numerical sample integration requested, but no levels specified.

• BUG: Variable scaling now works with multistart methods.

• BUG: DDACE and post=run needn’t require a seed; verified all post-run

Known Limitations