Version for WH10 Experiments

Submitted to /agh/git/kermor repository on 2011-06-20 Branch name: kermor04

(Daniel Wirtz, 2011-05-19) Disconnected the Approx classes from taking a BaseModel instance at approx computation. This way external tools can use the approximation algorithms, too.

Class approx.KernelApprox

(Daniel Wirtz, 2011-05-19) Disconnected the Approx classes from taking a BaseModel instance at approx computation. This way external tools can use the approximation algorithms, too.

(Daniel Wirtz, 2011-05-03) Removed off-property dependencies in computeCoeffs routines as expansion offsets are no longer used.

Class approx.TPWLApprox

(Daniel Wirtz, 2011-05-10)

The \(A_i\) matrices are now sparse if applicable (numel ~= 0 / numel < .5)
Removed the WeightFun property and reintroduced the Beta property, correctly forwarding the setting to the internally used gauss kernel.

Class Devel

(Syed Ammar, 2011-06-02) Added comments to the function setup

Class error.BaseCompLemmaEstimator

(Daniel Wirtz, 2011-05-29) Changed this classes name to "BaseKernelEstimator".

Class error.BaseEstimator

(Daniel Wirtz, 2011-05-23) Created a new interface with separate output error computation. Postprocessing is now a template method and a new property error.BaseEstimator.OutputError has been introduced. Renamed the LastError property to error.BaseEstimator.StateError.

Class error.DefaultEstimator

(Daniel Wirtz, 2011-06-01) Introduced a new property error.DefaultEstimator.RealOutputError to reflect that the actual output error also cannot be obtained by this error estimator. See the class comments for details.

(Daniel Wirtz, 2011-05-23) Adopted to the new error.BaseEstimator interface with separate output error computation.

Class error.GLEstimator

(Daniel Wirtz, 2011-05-29) Restructured the error estimators to better adopt to the current formulation. Now the KernelEstimators have a function getBeta instead of implementing the evalODEPart by themselves.

(Daniel Wirtz, 2011-05-25) Changed the implementation to correspond to the new comparison lemma estimator derivation. This requires only one ODE dimension and is a sharper estimate than before.

(Daniel Wirtz, 2011-05-23) Adopted to the new error.BaseEstimator interface with separate output error computation.

Class error.IterationCompLemmaEstimator

(Daniel Wirtz, 2011-05-29)

Changed this classes name to "LocalKernelEstimator".
Restructured the error estimators to better adopt to the current formulation. Now the KernelEstimators have a function getBeta instead of implementing the evalODEPart by themselves.

(Daniel Wirtz, 2011-05-25) Changed the computations to the comparison lemma type. This reduced the needed extra ODE dimensions to one and speeds up the evaluation process.

(Daniel Wirtz, 2011-05-23) Adopted to the new error.BaseEstimator interface with separate output error computation.

Class error.lipfun.ImprovedLocalSecantLipschitz

(Daniel Wirtz, 2011-06-07) Moved the ModifiedNewton methods from this class to kernels.BellFunction as they are more appropriate at that place.

Class error.lipfun.LocalGradientLipschitz

(Daniel Wirtz, 2011-05-29) Changed \(x_0\) to \(r_0\) to adopt new notation.

Class EstimatorAnalyzer

(Daniel Wirtz, 2011-05-20) Adopted to the new strategy pattern implemented for the LocalLipschitzFcn inside the LocalLipschitzErrorEstimator (now having a class instead of a function handle).

Class general.AffParamMatrix

(Syed Ammar, 2011-05-06) Implemented Setters for the class properties

Class general.interpolation.KernelInterpol

(Daniel Wirtz, 2011-05-03) Removed the artificial offset term \(b\) from the interpolation process (no longer used in kernel expansions)

Class general.POD

(Syed Ammar, 2011-05-06) Implemented Setter for UseSVDS flag

Class general.regression.BaseScalarSVR

(Daniel Wirtz, 2011-05-31) Removed the svidx parameter from the regress interface.

Class general.regression.KernelLS

(Daniel Wirtz, 2011-05-03) Removed \(b\) offset terms as no longer used in kernel expansions.

Class general.regression.ScalarEpsSVR

(Daniel Wirtz, 2011-06-01) Fitted the interface to the changed one in general.regression.BaseScalarSVR<p> (Syed Ammar, 2011-05-06) Implemented Setter for the property eps

(Daniel Wirtz, 2011-05-03) Removed the \(b\) offset terms from the SVM formulation (i.e. removing the equality constraint for the coefficient vectors) as kernel expansions no longer use offset terms.

Class general.regression.ScalarNuSVR

(Daniel Wirtz, 2011-06-01) Fitted the interface to the changed one in general.regression.BaseScalarSVR<p> (Daniel Wirtz, 2011-05-04) Removed offset \(b\) terms from computations.

Class KerMor

(Daniel Wirtz, 2011-05-03)

Committed KerMor version 0.3 to GIT.
Different documentation creation subroutines for unix and windows now, windows not supported yet
Version Number now gets passed to the documentation creation script

Class kernels.BellFunction

(Daniel Wirtz, 2011-06-07) Moved the ModifiedNewton methods from error.lipfun.ImprovedLocalSecantLipschitz to this class as they are more appropriate here.

(Daniel Wirtz, 2011-05-27) Changed \(x_0\) to \(r_0\) and \(x_R\) to \(r_m\) as adopted in the WH10 Paper.

(Daniel Wirtz, 2011-05-20) Removed the getLipschitzFunction method as it causes LARGE overhead when being called very often. Instead, the estimator always uses the improved estimation procedure.

(Daniel Wirtz, 2011-05-19)

Removed the PenaltyFactor property as no longer needed.
Modified the newton iteration penalization such that the 2nd degree polynomials extend the objective function both in value and derivative. This ensures propert continuation of the objective function into the penalized area with respect to the current kernel configuration (i.e. kernels.GaussKernel: small Gamma property causes high derivatives and thus large estimations.

Class kernels.KernelExpansion

(Daniel Wirtz, 2011-05-03)

Removed the off property as proper RKHS functions dont have an offset
No more setter for Ma property - too slow even during offline phase

Class models.BaseFirstOrderSystem

(Daniel Wirtz, 2011-05-29) Re-changed the \(x0(\mu)\) property back to be a function handle. Higher flexibility during subsequent simulations are to be rated higher than proper setting of the property, which should be assumed is being done (especially) for \(x0(\mu)\)

(Daniel Wirtz, 2011-05-13) Created new transient properties BaseFirstOrderSystem.mu and BaseFirstOrderSystem.u in order to store the current \(\mu\) and \(u(t)\) used for simulations on a higher level. Removed the old getODEFun function and replaced it by a direct implementation ODEFun. This enables to avoid nested function handles with in turn allow for a speedup of reduced simulations by almost a factor of 2.

Class models.BaseFullModel

(Syed Ammar, 2011-05-11) Implemented setters for the preApproximationTrainingCallback and postApproximationTrainingCallback

Class models.ReducedSystem

(Daniel Wirtz, 2011-05-13) Removed the old getODEFun function and replaced it by a direct implementation ODEFun. This enables to avoid nested function handles with in turn allow for a speedup of reduced simulations by almost a factor of 2.

Class solvers.ExplEuler

(Daniel Wirtz, 2011-05-31) Added a new middle class solvers.BaseCustomSolver which extracts the getCompTimes into a new abstraction layer.

Class solvers.Heun

(Daniel Wirtz, 2011-05-31) Added a new middle class solvers.BaseCustomSolver which extracts the getCompTimes into a new abstraction layer.

Member Utils.saveFigure (fig, filename, ext)

(Daniel Wirtz, 2011-05-31) Improved the export capabilites and automatic removement of any figure margins is performed.