The SUMO toolbox  2018a
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Pages
Todo List
Member AggregateObjective::scoreCandidates (var points, var state)
aggregate derivatives
Class BasicGaussianProcess

Refactor correlation functions into proper basis function class hierarchy.

solve the correlation matrix vs covariance matrix issue

Member BasicGaussianProcess::imse ()
Implement generic monte carlo integration
Member BasicGaussianProcess::marginalLikelihood (var dpsi, var dsigma2)
Adjoint derivatives work, but are very slow due to naive implementation
Member BasicGaussianProcess::updateRegression (var F, var hp)
Rho is only used by co-kriging, can we abstract this somehow ?
Class CandidateRanker
dscores (derivatives) are not scaled in CandidateRanker.score
Class CoKriging
Generalize to an arbitrary number of (multi-fidelity) datasets
Class CombinedModelBuilder

models are scored twice now

best models are plotted (good) but the best model can be worse than the best model of another modelbuilder

Member Constraint::evaluate (var x)

remove preEvaluation if transformation func's work properly

transformation functions don't handle input selection, dummy, default values, etc...

transformation functions don't handle input selection, dummy, default values, etc...

transformation functions don't handle input selection, dummy, default values, etc...

transformation functions don't handle input selection, dummy, default values, etc...

Member Constraint::initNewSamples (var state)
copied from CandidateRanker, should be more general (just init(state))
Member DiscreteOptimizer::DiscreteOptimizer (var varargin)

if input not discrete

discrete parameters assumed to be natural numbers ideally we can get a list of possible values each discrete parameter can take Simulation script can scale from 1:n to whatever they need

Member EnsembleModel::optimizeWeights ()
use already calculated model score for the weights ?
Member expectedImprovementEuclidean::scoreCandidates (var points, var state)

let metamodel decide what it can handle ?

doesnt work for model with multiple outputs (need for-loop)

mosbo EI derivatives

Member expectedImprovementHypervolume::scoreCandidates (var points, var state)
let metamodel decide what it can handle ?
Class ExtinctionPrevention
Code is hard to understand and could use cleaning up
Member findMinimum (var samples, var values)
add constraint handling back in it is removed for now as the Singleton doesn't play nice
Member instantiateClassOrFunction (var node, var parentConfig, var defaultType)
problem if type is not a class and defaultType is an abstract class
Member Kriging::getDefaultOptions ()
Apparantly Matlab 2008b doesn't auto forward the (static) call to BasicGaussianProcess::getDefaultOptions()
Member KrigingFactory::crossover (var parents, var options, var nvars, var FitnessFcn, var unused, var thisPopulation)
mix basisfunctions with different number of hyperparameters
Member KrigingFactory::KrigingFactory (var varargin)
this can be changed to IPS once we use proper classes for the BasisFunctions
Member LSSVMModel::constructInModelSpace (var samples, var values)
construct class in CTor and use model = changelssvm(model,'gam',1.2); to set the parameters + training data here
Member MaximinDistanceConstraint::getInternal ()
what's internal representation ?
Member MaximinDistanceConstraint::MaximinDistanceConstraint (var config)
fixed to 2D now
Member Model::basicPlotModel ()
not sure why this would be needed
Member Model::evaluateBatch (var points, var batchSize)
some duplicate code, see evaluateInModelSpaceBatch
Member NonlinearConstraint::getInternal ()
put transform in between ?
Member OptimizeCriterion::selectSamples (var state)
outDim is dependent on criterion, not state.values...
Class ParetoFront
ideas to improve the speed: update the front on new samples instead of recalculating everything, do a binarysearch (per dimension, not octree or branch&bound-like), this comes close to classification (finding the boundary where condition test 1 or 0) -> perhaps this can be solved using an SVM classifier ? :-)
Member ParetoFront::ParetoFront (var varargin)
new paretofront metric (convergence measure)
Member ParetoFront::updateParetoFront (var values)
debug plot of the pareto front. If more components are going to use it then it should be put in SUMO::runSampling (or samplemanager::add) but for now, all multiobjective methods use the ParetoFront class
Member PipelineSequentialDesign::selectSamples (var state)
newsamples isnt sorted!
Member plotParetoFront (var scores, var opts)
plot otherScores, but clip the values within pareto front range (for dtlz1 problem)
Member PolynomialFactory::getModelType ()
only this function is implemented from GeneticFactory
Member probabilityOfFeasibility::scoreCandidates (var points, var state)

doesnt work for model with multiple outputs (need for-loop)

calculation of PoF derivatives

Member probabilityOfImprovement::scoreCandidates (var points, var state)

doesnt work for model with multiple outputs (need for-loop)

test 2D output for regression (paretoFront is no longer sorted)

Member SampleManager::add (var newSampesUnfiltered, var newValuesUnfiltered, var newSampleIds)

Dummy values don't work at the moment:

  • samples -> prepareForEvaluation -> might give duplicates that are evaluated twice. If we remove dups before evaluation we should keep track of the removed ones... too much work
  • add: removeDups should be done on M3 space samples, as simulator space samples are thrown away which are in fact unique samples (with their dummy values). too much work conclusion: dummy values don't work

multiobjective case -> keep record of best pareto front ?

Member sliceContourPlot (var h, var object, var opts)

for 3D/5D problem make optimal use of y-axis

generalize nrSlices calculation

generalize sorting (up to 6D at least)

nrSlices per dim

prettify

plot fixedSamples

Member ValidationSet::ValidationSet (var varargin)
ValidationSet works only inside SUMO because of SampleManager