Version 5.0 (12/21/09)

Highlights

  • DAKOTA changes license to the GNU Lesser General Public License to enable library mode users, incorporating license changes for several TPLs
  • All new JAGUAR 2.0 graphical user interface for creating, editing, and running DAKOTA input files (BSD-like license)
  • Additional discrete range and discrete set types within design, uncertain, and state variables. New types supported by parameter studies, nondeterministic sampling, JEGA, and COLINY.
  • Anisotropic sparse grids, numerically-generated orthogonal polynomials, and improved expansion tailoring for stochastic expansion UQ methods
  • New methods for epistemic and mixed aleatory-epistemic uncertainty quantification: local/global interval estimation and local/global evidence.
  • Creation and management of evaluation working directories
  • SNL: DAKOTA modules on compute clusters (module avail dakota)

Usability/Core Execution

  • JAGUAR 2.0 graphical user interface, based on Eclipse Workbench, with text- and graphical-based editing, templates, sensitivity analysis wizard, and error checking. Available via separate download after registering for DAKOTA download.
  • Capability to specify working directories, template directories, and lightweight linking for system and fork interfaces.
  • DAKOTA input files can use Matlab-style sequence notation (L:S:U and beyond) to specify variable ranges; improved tolerance for whitespace in input files
  • Improved DAKOTA + application parallelism examples; new asynchronous local evaluation static scheduling for improved parallel tiling on clusters.
  • Pre-run (with optional variables output) and post-run (with variables/responses input) modes supported by sampling, paramter study, and DACE methods
  • Examples of Matlab and Python interfaces
  • Automated cleanup of DAKOTA temporary files on unexpected exit

Variables

  • Refactor of Variables and Constraints hierarchies to manage continuous, discrete integer, and discrete real domains among design, aleatory uncertain, epistemic uncertain, and state types.
    • Additional discrete design variable types: discrete design integer range, discrete design integer set, discrete design real set
    • New discrete uncertain distibutions: poisson, binomial, negative binomial, geometric, hypergeometric
    • Additional discrete state variable types: discrete state integer range, discrete state integer set, discrete state real set
  • Updates to specification of continuous and discrete histograms
  • Alternate specification for lognormal using lambdas and zetas

Methods (general)

  • New capability to perform parameter studies and sampling over discrete variables. New discrete variable types also supported in JEGA and COLINY.
  • DACE and parameter study classes (DDACE, FSUDACE, and multi-dimensional parameter study) can have correlations calculated and printed in addition to sampling methods
  • Improved accuracy and robustness in correlation computations
  • PSUADE: use updated (Campolongo 2007) sample generation scheme to improve space-filling properties
  • Improved EGO convergence controls (based on nearest neighbor)
  • Beta capabilities:
    • wrapper class for LANL's GPMSA code (Bayesian calibration)
    • importance sampling capability
    • nonlinear conjugate gradient optimization solver using Trilinos vector-matrix utilities
  • Bug fixes:
    • hang in DDACE orthogonal array LHS
    • OPT++ NIP methods not respecting bound constraints
  • New example problems for multifidelity OUU (MVFOSM as low fidelity UQ and stochastic expansion as high fidelity UQ), epistemic UQ, mixed aleatory-epistemic UQ, discrete sampling and parameter studies, etc.

Uncertainty Quantification (UQ) Capabilities:

  • Stochastic expansions (polynomial chaos expansion (PCE) and stochastic collocation (SC)):
    • Simplified controls for PCE: expansion formulation now inferred or estimated from quadrature_order or sparse_grid_level specifications. Automatic expansion tailoring minimizes performance loss due to poor expansion/integration synchronization.
    • Addition of numerically-generated orthogonal polynomials for generating optimal basis for non-Askey distribution types (uses Gauss-Wigert or discretized Stieltjes procedures for polynomial recursion coefficients in combination with Golub-Welsch for Gauss point/weight computation)
    • Addition of analytic variance-based decomposition for global sensitivity analysis using method of Sobol indices
    • Addition of analytic covariance among multiple response functions.
    • Addition of anisotropic Smolyak sparse grids, with user-supplied dimension preference.
    • Revision of Gaussian quadrature rules within sparse grids to use linear growth, providing finer grain control and more uniform integrand coverage.
    • PCE customizations for sparse grids: exclusive usage of linear growth rules (no Clenshaw-Curtis or Gauss-Patterson), total-order expansions for isotropic, custom expansion for anisotropic.
    • Addition of Askey and Wiener basis polynomial over-rides.
    • Modified variable correlation logic to revert to Wiener expansions for de-correlation as needed on a per-variable basis.
    • all_variables mode now includes epistemic uncertain variables (previously design and state only) using a Legendre basis.
  • Epistemic UQ methods:
    • Refactor of NonDInterval hierarchy to generalize and incorporate new methods.
      • Evidence: new global (LHS or EGO) and local (SQP or NIP) methods to calculate belief and plausibility in evidence theory calculations. Removed dependence on legacy Fortran package.
      • Interval estimation: new global (LHS or EGO) and local (SQP or NIP) methods to calculate response output intervals.
    • Extension of active Variables/Constraints views to support aleatory, epistemic, or aggregated aleatory-epistemic uncertain views.
  • Mixed aleatory-epistemic UQ methods:
    • New nested approaches (e.g. second-order probability, mixed evidence) where the outer loop (e.g. interval estimation or evidence) can leverage analytic moments and their sensitivities with respect to epistemic parameters.

Framework Enhancements:

  • Eliminated dependence on GSL in favor of Boost (to eliminate GNU GPL dependency)
  • Fully deployed Teuchos for all numerical data types. Boost and STL are used for all bookkeeping types.
  • Better out-of-source (VPATH) builds and documentation, including management of Boost and Teuchos
  • Switched from RNUM2 to Boost MT19937 random number generator for longer period, with run-time selection option
  • Deployed Boost multi-index containers to evaluation queue management, supporting multiple indexing options for optimized lookups.
  • Lightweight active/inactive data views implemented with Teuchos and Boost multi-array to eliminate deep data copies.
  • Updates to third-party libraries: JEGA, PSUADE, AMPL, Boost, Pecos, Teuchos
  • Minimal data mode for variables and responses omits labels, types, ids in some contexts for better scalability
  • Unified bounds checking for debugging and reduced overhead in optimized builds

Miscellaneous:

  • Binary distributions for Mac OS X and Cygwin, including library dependencies
  • Better detection of MPICH2 and shmem communicators
  • Automatic syncronization of DAKOTA input specification help docs to JAGUAR
  • Surfpack: improved random seed management for repeatability.
  • Better handling of numerical comparisons on 32-bit
  • Finite differencing honors variable bounds; more efficient finite difference Hessians
  • PSUADE generally available under LGPL
  • Surfpack and LHS relicensed under LGPL
  • DLL API improvements for re-entry, error handling, get/set options
  • ModelCenter, MATLAB, and Python interface updates to latest API
  • Updates to Matlab memory management
  • HTTP-based usage tracking (disabled by default; currently used Sandia internally only)
  • Windows: use spawn as alternative to fork