history

Version 1.7 (2015-09-28)
-----------

* Methods

- Added the nudging method, implemented in 'Nudging'.
- Added 'FrontPositionObserver', designed to work with
  'LevelSetObservationManager'.

* Models

- Added two new methods to 'QuadraticModel': 'GetStateErrorVarianceProjector()'
  and 'GetStateErrorVarianceReduced()'.
- Added a new method to 'QuadraticModel': 'SetMPICommunicator()'.

* Observation Manager

- Added the level-set observation manager 'LevelSetObservationManager'.
- Added a new method to 'LinearObservationManager' and
  'ObservationManagerTemplate': 'GetNudgingMatrix'.

* Tools

- Added a test to check the consistency between reduced-order extended Kalman
  filter and extended Kalman filter.
- Added a test to check the consistency between the nudging and the extended
  Kalman filter.
- Added a new test to the adjoint model test case.

* Interface

- Renamed the 'Forward()' method to 'Prediction()'. Added a method 'Forward()'
  which regroups 'Prediction()' and 'Analyze()'.
- Modified the file structure of Verdandi to compile Verdandi as a
  library. The non-template files have been separated from the template ones
  in the "share" directory.

* Documentation

- Improved the documentation about the installation under Linux and Windows.
- Added documentation about the nudging.
- Added documentation about the level-set observation manager.
- Improved the tutorial about the installation of Google test.

* Bug fixes

- In the reduced-order unscented Kalman filter, a matrix was updated at each
  step while it should be updated only at analysis steps
- ROUKF was unable to analyze the first step.
- The binary flag mode was missing when opening input and output files.
- An include was missing under MacOs.
- A variable declaration in "blue.hpp" did not compile with gcc 5.2.
- Changed the system includes for the compilation under Windows.
- A template was not correctly instantiated in the 'ClampedBar' model.

* Miscellaneous

- Improved the tests: all comparisons between methods are using the same test set up.
- Added 'MessageHandler::Recipient_map' to provide access to the static
  variable 'recipient_map'.


Version 1.6.1 (2014-11-21)
-------------

* Bug fixes

- In "PythonModel.hxx" and "PythonObservationManager.hxx", the macro providing
  the Python version was not working with Clang.
- Various examples are now using the 'RandomPerturbationManager' by default
  when compiling with Clang.
- 'ShallowWater' uses 'Newran' by default when compiling with Clang.

* Miscellaneous

- When compiling with Clang, SCons will automatically add the C++0x option.


Version 1.6 (2014-10-29)
-----------

* Methods

- Added support for model error in 'ExtendedKalmanFilter'.
- Improved the display options of 'ExtendedKalmanFilter'.
- Added the option to read the perturbations of 'MonteCarlo' in files instead
  of generating them randomly.
- Removed useless copies in 'MonteCarlo'.
- Added the option to change the initial weight in
  'DiscountedRidgeRegression'.
- In all methods, the initial condition is now saved before assimilation and
  even in the case the model is not initialized by the method.
- Initial import of 'ExtendedMinimaxFilter'.

* Tools

- Added a new perturbation manager: 'RandomPerturbationManager', using the
  C++11 library <random>.
- Newran is now optional everywhere.
- Rewrote CPPUnit tests into the framework of Google Test.
- Added tests for 'Model::SetTime', the adjoint model and the tangent linear
  operator of the model.
- Added a test to check that the unscented Kalman filter and the extended
  Kalman filter give the same results with a linear model and the same
  observations.
- Added the directory "test/unit/tool/", primarily with tests on the
  perturbation managers.
- Added the new module "module_performance" aiming at keeping track of the
  time inside tests.
- Made use of the perturbation managers in the unit tests.

* Interface

- Updated the Python interface to be consistent with the expected description
  of uncertain parameters.
- Added the option not to call the Python method 'Initialize' from
  'PythonModel::Initialize'.
- Added 'PythonModel::GetModelInstance()'.

* Documentation

- Improved the documentation about the installation under Linux and Windows.
- Added documentation about unit tests.

* Bug fixes

- In 'MonteCarlo::InitializeStep', the model method 'ParameterUpdated' was
  called for parameters with option "init_step".
- In 'TR1PerturbationManager', the normal sampling could generate the same
  values when the corresponding calls were too close (in time).
- The arguments 'initialize_model' and 'initialize_perturbation_manager' were
  not transferred from 'MonteCarlo::Initialize(string, bool, bool)' to
  'MonteCarlo::Initialize(VerdandiOps&, bool, bool)'.
- In sequential mode, 'ReducedOrderUnscentedKalmanFilter' saved the analyses
  with the forecasts.
- The name of the logger output file was set after the first strings were sent
  to the logger.
- The output saver did not append a new line after the data are written in
  'WriteText'.
- The initialization and finalization of MPI were missing in
  'ForwardDriver.cxx'.
- Moved a call to the logger that was made before the model was initialized.
- Moved the definitions of the default parameters from "Error.cxx" to
  "Error.hxx", which avoids compilation failure with given compilers
  (including Clang).
- 'FourDimensionalVariational' could not compile without the trajectory
  manager.
- The copy constructor of 'TR1PerturbationManager' did not duplicate the
  underlying random number generators.

* Miscellaneous

- Renamed "display.show_iteration" and "display.show_time" to
  "display.iteration" and "display.time" in the configurations.
- Improved the log messages and made them consistent across methods.
- Improved the detection of C++11.
- Added the option to use either TR1 or Newran in the shallow water model.
- Removed an unused function in observation managers: 'bool
  HasObservation(double time)'.


Version 1.5 (2013-02-28)
-----------

* Methods

- In 'FourDimensionalVariational', added the possibility to add a term to the
  cost function and to the adjoint state by the use of two model functions:
  'GetAdditionalAdjointTerm()' and 'GetAdditionalCostTerm()'.

* Documentation

- Added documentation about plugging an observation manager.
- Added a PDF version of the documentation.
- Improved the introduction to the Python interface.
- Updated the documentation for the use of Numpy in 'PythonModel'.

* Bug fixes

- In 'LinearObservationManager.py', the parameter 'Nskip' was not used.
- In the Python implementation of 'QuadraticModel', the function
  'ApplyTangentLinearOperator' was erroneous when the model was defined with
  both the quadratic term and the linear term.
- The fluxes in the shallow-water model were not properly computed. The errors
  had little consequences only in case the space steps and the height were
  close to 1.
- Verdandi directory "include/" was not given to g++ in
  "bin/methods_requirement".

* Miscellaneous

- Added 'OutputSaver::IsSaved(string variable_name)'.
- Cleaned up "bin/methods_requirement".


Version 1.4 (2012-11-13)
-----------

* Methods

- Removed the template parameter T in all assimilation methods.
- Added support for elementwise access to the tangent linear observation
  operator in 'EnsembleKalmanFilter' and 'FourDimensionalVariational'.
- Added the option to save the forecast and analysis times.
- Renamed the saved variables "state_forecast" and "state_analysis" to 
  "forecast_state" and "analysis_state".
- Improved 'UnscentedKalmanFilter' and 'ReducedOrderUnscentedKalmanFilter'
  with more suitable types for several intermediate variables.
- Modified 'HamiltonJacobiBellman' so that it can save the time in addition to
  the time step, and so that the output saver decides to save variables using
  the time, not the time step.
- Removed the deprecated calls to MPI in the form 'MPI::Function', instead of
  'MPI_Function'.
- Added tests to check whether MPI has been initialized/finalized before
  calling 'MPI_Init'/'MPI_Finalize'.

* Models

- Improved the model interface for the uncertain parameters.
- Initial import of 'CheckingModel'.
- Simplified 'ApplyOperator' to remove the option 'forward' which is now set
  to 'false' all time.

* Output saver

- Added the option to read and store observations in HDF5 format. Implemented
  an example for 'QuadraticModel'.

* Example models

- Added the forecast and analysis times in output variables.
- In 'QuadraticModel', added tests on the dimensions of uncertainty
  descriptions provided in the configuration.
- In 'ShallowWater', the state vector and the full state vector are now stored
  in two distinct attributes ('state_' and 'full_state_').

* Documentation

- Added the script "methods_requirements" in "bin/" to list the member
  functions of a model and of an observation manager that are required by a
  specified assimilation method.
- Improved the section "Plugging an Existing Model".
- Renamed the section "Using Verdandi" to "Dependencies" and improved its
  contents.
- Added documentation about the compilation options in "Debugging".
- Added a section in "Plugging a Model" for plugging a Python model.

* Bug fixes

- In 'EnsembleKalmanFilter', the empirical variance derived from
  the ensemble was (Nmember - 1) times higher than expected, where Nmember is
  the number of members in the ensemble.
- The example observation managers returned the wrong observations when the
  variable 'Nskip' was different from 1.
- In 'FourDimensionalVariational', the method did not compile when the option
  'VERDANDI_WITH_TRAJECTORY_MANAGER' was deactivated.
- In 'HamiltonJacobiBellman', the model was not initialized when the source
  term was included but not the advection term. 
- In 'UnscentedKalmanFilter', the sigma-point matrix was no longer updated.
- In the model 'ShallowWater', the size of the full state vector was erroneous.
- 'QuadraticModel<T>' was compiling only if 'T' was double.
- In 'HamiltonJacobiBellman', the model methods 'FinalizeStep' and 'Finalize'
  are now called only if the advection term or the source term are included.

* Miscellaneous

- Optimized the general SConstruct file: the link flags for the external
  libraries were set twice.
- Removed all exception specifications in class 'Error' and in its derived
  classes.


Version 1.3 (2012-06-27)
-----------

* Methods

- Modified the model interface so that it only returns references to
  large vectors and matrices.
  The model methods
  'GetTangentLinearOperator(tangent_linear_operator&)',
  'GetAdjointState(state&)', 'SetAdjointState(state&)',
  'GetStateErrorVarianceRow(int, state_error_variance_row&)' and
  'GetStateErrorVarianceSqrt(state_error_variance&, state_error_variance&)'
  have been replaced with respectively:
  'tangent_linear_operator& GetTangentLinearOperator()',
  'state& GetAdjointState()', 'AdjointStateUpdated()',
  'state_error_variance_row& GetStateErrorVarianceRow(int)',
  'state_error_variance& GetStateErrorVarianceProjector()' and
  'state_error_variance_reduced& GetStateErrorVarianceReduced()'.
- In 'ReducedOrderUnscentedKalmanFilter', the algorithm is now able to 
  parallelize several model tasks which in turn can be run in parallel.
- In the model interface, 'ApplyOperator' and 'ApplyTangentLinearOperator'
  now return the time associated with the computed state.
- Added an option to 'OptimalInterpolation' to compute the diagonal
  of the analysis error variance.
- Added an example for 'HamiltonJacobiBellman' with the quadratic model.

* Observation manager

- Modified the observation manager interface so that it only returns
  references to large vectors and matrices.
  The observation manager methods
  'GetObservation(observation&)', 'GetInnovation(state&, observation&)' and
  'GetTangentLinearOperatorRow(int, tangent_linear_operator_row&)'
  have been replaced with respectively:
  'observation& GetObservation()', observation& GetInnovation(state&)' and
  'tangent_linear_operator_row& GetTangentLinearOperatorRow(int)'.

* Example models

- Added the ability in 'PetscClamedBar' to define a MPI communicator different
  from MPI_COMM_WORLD.
- Improved the interface of 'QuadraticModel' for the uncertain parameters.

* Example observation managers

- Initial import of 'ObservationGenerator'.
- Improved the performance of 'PetscLinearObservationManager'.
- Added the ability in 'PetscLinearObservationManager' to define a MPI
  communicator different from MPI_COMM_WORLD.
- Modified 'GridToNetworkObservationManager' so that it can read
  observations stored in a Seldon vector format.

* Documentation

- Added documentation about the optimization solver.
- Improved documentation related to exceptions.


Version 1.2.1 (2012-04-10)
-------------

* Methods

- Added an option to 'ReducedOrderUnscentedKalmanFilter' to manage
  observations provided as innovations.

* Bug fixes

- Removed illegal 'typename' keywords in 'ModelTemplate' definitions.


Version 1.2 (2012-03-30)
-----------

* Methods

- The state vectors are now passed by reference to the assimilation methods.
  'GetState(state&)', 'SetState(state&)', 'GetFullState(state&)' and
  'SetFullState(state&)' have been replaced with 'state& GetState()',
  'StateUpdated()', 'state& GetFullState()' and 'FullStateUpdated()'.
  Also replaced the model methods 'GetStateLowerBound(state&)' and
  'GetStateUpperBound(state&)' with 'state& GetStateLowerBound()' and
  'state& GetStateUpperBound()'.
- 'ForwardDriver' and 'ReducedOrderUnscentedKalmanFilter' can now efficiently
  handle models parallelized by MPI.
- Initial import of sequential aggregation (for ensemble forecasting) with
  discounted ridge regression ('DiscountedRidgeRegression').
- Added support for parallel computations in 'EnsembleKalmanFilter'.
- Added support for parallel computations in 'ComputeBLUE_vector'.
- Enforced the rule that the member functions of the data assimilation
  methods are non const, except 'GetName'.
- Modified the interface to 'MonteCarlo' so that it is consistent with the
  interface to 'EnsembleKalmanFilter'.
- Made 'MonteCarlo' template of the perturbation manager.
- Added methods 'GetModel', 'GetObservationManager', 'GetPerturbationManager',
  'GetOuputSaver', 'FinalizeStep' and 'Finalize' in all appropriate
  assimilation methods.
- Removed the references to 'VERDANDI_DENSE' and 'VERDANDI_SPARSE' in
  'OptimalInterpolation'.
- Added consistency checks for the dimensions of the arguments of
  'ComputeBLUE_matrix'.

* Models

- Added the possibility to use a model written in Python. A generic interface
  between Python and C++ is implemented in 'PythonModel'.

* Observation managers

- Added the possibility to use an observation manager written in Python.  A
  generic interface between Python and C++ is implemented in
  'PythonObservationManager'.

* Example models

- Added the new model 'PetscClampedBar', which implements 'ClampedBar' with
  PETSc.
- Extended the interface of the quadratic model for the ensemble Kalman
  filter.
- Added configuration for the 'MonteCarlo' example with the quadratic model.

* Example observation managers

- Added the new observation manager 'PetscLinearObservationManager', which
  implements 'LinearObservationManager' with PETSc.
- Added support for non constant time intervals between observations in
  'LinearObservationManager'.

* Output saver

- Added 'OutputSaver::Activate' and 'OutputSaver::Deactivate'.

* Perturbation managers

- Added an interface to the random number generator library from C++ TR1.
- Set TR1 as the default random number generator library.

* Debugging

- Added the option "dump_env" to dump the SCons environment in the file
  "scons_env.log"

* Documentation

- Added the section "Plugging a Model".
- Added a description of the Lorenz model.

* Bug fixes

- In 'EnsembleKalmanFilter', from one member to another, the
  parameters were not perturbed around the same reference values.
- The examples in "example/template" did not compile.

* Miscellaneous

- Updated Seldon version from 5.1 (patched for Verdandi) to 5.2 (release
  candidate).
- Renamed 'TRY' to 'VERDANDI_TRY' and 'END' to 'VERDANDI_END'
- Initial import of "Functions_Vector2.*xx" and "Fuctions_Vector3.*xx".
- Added the error class 'ErrorPythonUndefined'.
- Moved "OutputSaver.*xx" and "Variable.*xx" to "share/".


Version 1.1 (2011-10-25)
-----------

* Methods

- Added an implementation of the ensemble Kalman filter
  ('EnsembleKalmanFilter').
- Modified 'UnscentedKalmanFilter' to support a model state of type
  'Vector<Collection>'.
- Improved the saving options in 'ReducedMinimax'.
- Added in 'MonteCarlo::Initialize' the option to choose to initialize or not
  the model and the perturbation manager.

* Example models

- Extended the interface of the shallow water model for the ensemble Kalman
  filter.

* Perturbation managers

- The perturbation manager based on Newran can now accept several types of
  seed, and it requires to explicitly choose this type in the configuration.

* Bug fixes

- In the reduced minimax filter, the observations were assimilated one model
  time step earlier than expected.


Version 1.0.2 (2011-08-23)
-------------

* Documentation

- Added documentation about the compilation of the SWIG interface under
  Windows.
- Updated the documentation for the Python interface with the current naming
  conventions.

* Improvements

- Made the SWIG interface compatible with Windows.

* Bug fixes

- In the perturbation managers, the mean and the covariance matrix were
  erroneously modified when the constraints were not satisfied by the first
  sampling.
- The homogeneous log-normal perturbations were erroneous.
- In the perturbation manager, several log-normal sampling functions did not
  work properly.


Version 1.0.1 (2011-06-16)
-------------

* Documentation

- Added a section in the documentation about the compilation of Verdandi for a
  64-bit architecture under Windows.

* Bug fixes

- Fixed compilation problems with Visual C++ compiler related to 'Ops::Ops'.
- 'ForwardDriver' ignored the configuration variables 'forward.output.log' and
  'forward.output.configuration'.
- Added "python" in the library list so that it may be linked when the SWIG
  interface is built. This is necessary under MacOS.
- The random number generators were allocated twice, which led to minor memory
  leaks.


Version 1.0 (2011-05-27)
-----------

* Methods

- Improved the handling of the sparse and dense options in the methods:
  EKF, UKF, ROEKF and ROUKF.

* Observations

* Example models

- Added a new method 'Model::GetNfull_state()'.
- Changed the default configuration for the shallow-water model example: the
  observations are now localized at specific points instead of all grid
  points.

* Documentation

- Initial import of the documentation for the perturbation manager.

* Miscellaneous

- Improved 'GetCholesky' in the sparse case: the 'RowSparse' matrix is no
  longer converted to a dense matrix but to an 'ArrayRowSymSparse' matrix
  before the decomposition.


Version 0.9 (2011-05-12)
-----------

* Methods

- Added an implementation of the 4D-Var method ('FourDimensionalVariational').
- Redesigned the perturbation manager so that adding new generators and new
  distributions is easier, and added support for the TRNG library.
- Improved the implementation of 'ExtendedKalmanFilter' and
  'ReducedOrderExtendedKalmanFilter'.
- Improved the consistency of 'MonteCarlo' with Verdandi.
- Initial import of the function 'ComputeBLUE_vector'.
- Added options to 'Method::Initialize' to indicate if the model or
  observation manager have to be initialized or not.
- Improved the snapshot recording so that $H^T y$ and $H^T H x$ can be
  recorded in 'ReducedMinimaxFilter'.
- Added support for a model error variance provided as a scaled identity
  matrix in 'ReducedMinimaxFilter'.
- Added the ability to save the square root of the model error variance
  in 'ReducedMinimaxFilter'.
- Changed the initialization of the assimilation methods: the constructor now
  takes no argument and the configuration file (as string or 'Ops' instance)
  is the argument of the method 'Initialize'.

* Observations

- Added the ability to define a sparse observation operator with a Lua table.

* Example models

- Added the uncertainty description for the constant term in 'QuadraticModel'.
- Removed 'ClampedBar' model, and renamed 'ParametricClampedBar' to
  'ClampedBar'.

* Documentation

- Largely improved the installation description.
- Initial import of the documentation for the example programs associated
  with the shallow-water model and the clamped-bar model.
- Added documentation on Lua configuration files.
- Added documentation for the Monte Carlo method.
- Made algorithms notation uniform in documentation.

* Miscellaneous

- Largely improved the compatibility with Visual C++.
- Added support for a user-defined dependency list in the "SConstruct".
- Extended the SWIG interface to the methods 'ForwardDriver' and
  'ReducedOrderExtendedKalmanFilter', and to the 'ClampedBar' model.
- Uniformized 'ClampedBar' example programs.

* Bug fixes

- In 'ReducedOrderExtendedKalmanFilter::Analyze()', the computation of the
  matrix 'U' was erroneous.
- The correlations between normal variables were properly handled only if they
  were equal to 0 or 1 in 'BasePerturbationManager'.
- The interpolation in 'ObservationAggregator' was erroneous.
- The tangent linear model was applied at time "h + 1"  in the method
  'PropagateCovianceMatrix()' of 'ExtendedKalmanFilter'.
- In 'ReducedOrderExtendedKalmanFilter::PropagateCovianceMatrix()', the model
  was applied at time "h + 1", but it was the tangent linear model that had
  to be applied at time "h".
- 'MessageHandler::RemoveRecipient' did not properly erase elements
  from the recipients map.
- The reduction failed when the number of snapshots was greater than
  the size of the state in 'ReducedMinimaxFilter'.
- An exception was raised when the size of the reduced state changed
  in 'ReducedMinimaxFilter'.
- In SWIG interface: the preprocessing directive 'VERDANDI_SPARSE' was defined
  in "verdandi.i" while the directive 'VERDANDI_DENSE' was defined
  in "verdandi.cpp".
- In the methods 'BasePerturbationManager::Sample' for vector collections,
  the perturbation were not deallocated.
- The time was not initialized in the constructors of 'ClampedBar'.
- Assembling Newmark matrices was erroneous in 'ClampedBar'.


Version 0.8 (2010-12-07)
-----------

* Methods

- Added the reduced-order extended Kalman filter
  ('ReducedOrderExtendedKalmanFilter').
- Added the reduced-order unscented Kalman filter
  ('ReducedOrderUnscentedKalmanFilter').
- Added the reduced minimax filter ('ReducedMinimax').
- Added support for Monte Carlo simulations ('MonteCarlo').
- Modified 'UnscentedKalmanFilter' to manage model states of type
  'VectorCollection'.

* Observations

- Improved the error management in 'LinearObservationManager'.
- Enabled reading a sparse observation operator from file in
  'LinearObservationManager'.

* Example models

- Added the model 'ParametricClampedBar'.

* Miscellaneous

- Added options to the SConstruct to provide Mumps and UMFPack compilation.
- Configured Ops so that, when an error occurs, the program throws an
  exception instead of aborting.
- Added option '%{rank}' to the Logger definition file to enable each process
  to have its own log file.

* Bug fixes

- The quadratic term was not properly computed in the method 'Forward' (and
  therefore in 'ApplyOperator' as well) of 'QuadraticModel'.
- The tangent linear model of 'QuadraticModel' was erroneous in case no
  quadratic term and no linear term were defined.


Version 0.7 (2010-09-05)
-----------

First tagged version.


The development started in 2008.