EfficientGlobalOptimization

class EfficientGlobalOptimization(*args)

Efficient Global Optimization algorithm.

The EGO algorithm [jones1998] is an adaptative optimization method based on kriging. An initial design of experiment is used to build a first metamodel. At each iteration a new point that maximizes a criterion is chosen as optimizer candidate. The criterion uses a tradeoff between the metamodel value and the conditional variance. Then the new point is evaluated using the original model and the metamodel is relearnt on the extended design of experiment.

Parameters:
problemOptimizationProblem

The optimization problem to solve

krigingResultKrigingResult

The result of the meta-model on the first design of experiment

noiseFunction

Noise model used on design experiment points, and also for improvement optimization

Notes

Each point added to the metamodel design seeks to improve the current minimum. We chose the point so as to maximize an improvement criterion based on the metamodel.

I(x_{new}) = max(f_{min} - Y_{new}, 0)

The default criteria is called EI (Expected Improvement) and aims at maximizing the mean improvement:

\mathbb{E}\left[I(x_{new})\right] = \mathbb{E}\left[max(f_{min} - Y_{new}, 0)\right]

This criterion is explicited using the kriging mean and variance:

\mathbb{E}\left[I(x_{new})\right] = (f_{min} - m_K(x_{new})) \Phi\left( \frac{f_{min} - m_K(x_{new})}{s_K(x_{new})} \right) + s_K(x_{new}) \phi\left( \frac{f_{min} - m_K(x_{new})}{s_K(x_{new})} \right)

An optional observation noise variance can be provided:

Y_{obs} = Y(x) + \sigma_{\epsilon}(x) \epsilon

In that case the AEI (Augmented Expected Improvement) formulation is used. As we don’t have access to the real minimum of the function anymore a quantile of the kriging prediction is used, with the constant c:

u(x) = m_K(x) + c s_K(x)

This criterion is minimized over the design points:

x_{min} = \argmax_{x_i} (u(x_i))

The AEI criterion reads:

AEI(x_{new}) = \mathbb{E}\left[max(m_K(x_{min}) - Y_{new}, 0)\right] \times \left(1 - \frac{\sigma_{\epsilon}(x_{new})}{\sqrt{\sigma_{\epsilon}^2(x_{new})+s^2_K(x_{new})}} \right)

with

\mathbb{E}\left[max(m_K(x_{min}) - Y_{new}, 0)\right] = (m_K(x_{min}) - m_K(x_{new})) \Phi\left( \frac{m_K(x_{min}) - m_K(x_{new})}{s_K(x_{new})} \right) + s_K(x_{new}) \phi\left( \frac{m_K(x_{min}) - m_K(x_{new})}{s_K(x_{new})} \right)

A less computationally expensive noise function can be provided through setNoiseModel() to evaluate \sigma^2_{\epsilon}(x) for the improvement criterion optimization, the objective being only used to compute values and associated noise at design points.

By default the criteria is minimized using MultiStart with starting points uniformly sampled in the optimization problem bounds, see setMultiStartExperimentSize() and setMultiStartNumber(). This behavior can be overridden by using another solver with setOptimizationAlgorithm().

Examples

>>> import openturns as ot
>>> ot.RandomGenerator.SetSeed(0)
>>> dim = 4
>>> model = ot.SymbolicFunction(['x1', 'x2', 'x3', 'x4'],
...     ['x1*x1+x2^3*x1+x3+x4'])
>>> model = ot.MemoizeFunction(model)
>>> bounds = ot.Interval([-5.0] * dim, [5.0] * dim)
>>> problem = ot.OptimizationProblem()
>>> problem.setObjective(model)
>>> problem.setBounds(bounds)
>>> experiment = ot.Composite([0.0] * dim, [1.0, 2.0, 4.0])
>>> inputSample = experiment.generate()
>>> outputSample = model(inputSample)
>>> print('Initial minimum output value: ', outputSample.getMin())
Initial minimum output value:  [-248]
>>> covarianceModel = ot.SquaredExponential([2.0] * dim, [0.1])
>>> basis = ot.ConstantBasisFactory(dim).build()
>>> kriging = ot.KrigingAlgorithm(inputSample, outputSample, covarianceModel, basis)
>>> kriging.run()
>>> algo = ot.EfficientGlobalOptimization(problem, kriging.getResult())
>>> algo.setMaximumCallsNumber(2)
>>> algo.run()
>>> result = algo.getResult()
>>> updatedKrigingResult = algo.getKrigingResult()
>>> updatedOutputSample = updatedKrigingResult.getOutputSample()
>>> yMin = updatedOutputSample.getMin()

Methods

getAEITradeoff()

AEI tradeoff constant accessor.

getCheckStatus()

Accessor to check status flag.

getClassName()

Accessor to the object's name.

getCorrelationLengthFactor()

Correlation length stopping criterion factor accessor.

getExpectedImprovement()

Expected improvement values.

getImprovementFactor()

Improvement criterion factor accessor.

getKrigingResult()

Retrieve the Kriging result.

getMaximumAbsoluteError()

Accessor to maximum allowed absolute error.

getMaximumCallsNumber()

Accessor to maximum allowed number of calls.

getMaximumConstraintError()

Accessor to maximum allowed constraint error.

getMaximumIterationNumber()

Accessor to maximum allowed number of iterations.

getMaximumRelativeError()

Accessor to maximum allowed relative error.

getMaximumResidualError()

Accessor to maximum allowed residual error.

getMaximumTimeDuration()

Accessor to the maximum duration.

getMetamodelNoise()

Metamodel noise model accessor.

getMultiStartExperimentSize()

Size of the design to draw starting points.

getMultiStartNumber()

Number of starting points for the criterion optimization.

getName()

Accessor to the object's name.

getNoiseModel()

Improvement noise model accessor.

getOptimizationAlgorithm()

Expected improvement solver accessor.

getParameterEstimationPeriod()

Parameter estimation period accessor.

getProblem()

Accessor to optimization problem.

getResult()

Accessor to optimization result.

getStartingPoint()

Accessor to starting point.

hasName()

Test if the object is named.

run()

Launch the optimization.

setAEITradeoff(c)

AEI tradeoff constant accessor.

setCheckStatus(checkStatus)

Accessor to check status flag.

setCorrelationLengthFactor(b)

Correlation length stopping criterion factor accessor.

setImprovementFactor(improvementFactor)

Improvement criterion factor accessor.

setMaximumAbsoluteError(maximumAbsoluteError)

Accessor to maximum allowed absolute error.

setMaximumCallsNumber(maximumCallsNumber)

Accessor to maximum allowed number of calls

setMaximumConstraintError(maximumConstraintError)

Accessor to maximum allowed constraint error.

setMaximumIterationNumber(maximumIterationNumber)

Accessor to maximum allowed number of iterations.

setMaximumRelativeError(maximumRelativeError)

Accessor to maximum allowed relative error.

setMaximumResidualError(maximumResidualError)

Accessor to maximum allowed residual error.

setMaximumTimeDuration(maximumTime)

Accessor to the maximum duration.

setMetamodelNoise(metaModelNoise)

Metamodel noise model accessor.

setMultiStartExperimentSize(...)

Size of the design to draw starting points.

setMultiStartNumber(multiStartNumberSize)

Number of starting points for the criterion optimization.

setName(name)

Accessor to the object's name.

setNoiseModel(noiseModel)

Improvement noise model accessor.

setOptimizationAlgorithm(solver)

Expected improvement solver accessor.

setParameterEstimationPeriod(...)

Parameter estimation period accessor.

setProblem(problem)

Accessor to optimization problem.

setProgressCallback(*args)

Set up a progress callback.

setResult(result)

Accessor to optimization result.

setStartingPoint(startingPoint)

Accessor to starting point.

setStopCallback(*args)

Set up a stop callback.

getMaximumEvaluationNumber

setMaximumEvaluationNumber

__init__(*args)
getAEITradeoff()

AEI tradeoff constant accessor.

Returns:
cfloat

Used to define a quantile of the kriging prediction at the design points. u(x)=m_K(x)+c*s_K(x)

getCheckStatus()

Accessor to check status flag.

Returns:
checkStatusbool

Whether to check the termination status. If set to False, run() will not throw an exception if the algorithm does not fully converge and will allow one to still find a feasible candidate.

getClassName()

Accessor to the object’s name.

Returns:
class_namestr

The object class name (object.__class__.__name__).

getCorrelationLengthFactor()

Correlation length stopping criterion factor accessor.

When a correlation length becomes smaller than the minimal distance between design point for a single component that means the model tends to be noisy, and the EGO formulation is not adapted anymore.

Returns:
bfloat

Used to define a stopping criterion on the minimum correlation length: \theta_i < \frac{\Delta_i^{min}}{b} with \Delta^{min} the minimum distance between design points.

getExpectedImprovement()

Expected improvement values.

Returns:
eiSample

The expected improvement optimal values.

getImprovementFactor()

Improvement criterion factor accessor.

Returns:
alphapositive float, default=0.0 (disables the criterion)

Used to define a stopping criterion on the improvement criterion: I_{max} < \alpha |Y_{min}| with I_{max} the current maximum of the improvement and Y_{min} the current optimum.

getKrigingResult()

Retrieve the Kriging result.

Returns:
krigingResultKrigingResult

Kriging result that takes all observations into account.

Notes

Before run() is called, this method returns the KrigingResult passed to the constructor. Once run() has been called, it returns an updated KrigingResult that takes new observations into account.

getMaximumAbsoluteError()

Accessor to maximum allowed absolute error.

Returns:
maximumAbsoluteErrorfloat

Maximum allowed absolute error, where the absolute error is defined by \epsilon^a_n=\|\vect{x}_{n+1}-\vect{x}_n\|_{\infty} where \vect{x}_{n+1} and \vect{x}_n are two consecutive approximations of the optimum.

getMaximumCallsNumber()

Accessor to maximum allowed number of calls.

Returns:
maximumEvaluationNumberint

Maximum allowed number of direct objective function calls through the () operator. Does not take into account eventual indirect calls through finite difference gradient calls.

getMaximumConstraintError()

Accessor to maximum allowed constraint error.

Returns:
maximumConstraintErrorfloat

Maximum allowed constraint error, where the constraint error is defined by \gamma_n=\|g(\vect{x}_n)\|_{\infty} where \vect{x}_n is the current approximation of the optimum and g is the function that gathers all the equality and inequality constraints (violated values only)

getMaximumIterationNumber()

Accessor to maximum allowed number of iterations.

Returns:
maximumIterationNumberint

Maximum allowed number of iterations.

getMaximumRelativeError()

Accessor to maximum allowed relative error.

Returns:
maximumRelativeErrorfloat

Maximum allowed relative error, where the relative error is defined by \epsilon^r_n=\epsilon^a_n/\|\vect{x}_{n+1}\|_{\infty} if \|\vect{x}_{n+1}\|_{\infty}\neq 0, else \epsilon^r_n=-1.

getMaximumResidualError()

Accessor to maximum allowed residual error.

Returns:
maximumResidualErrorfloat

Maximum allowed residual error, where the residual error is defined by \epsilon^r_n=\frac{\|f(\vect{x}_{n+1})-f(\vect{x}_{n})\|}{\|f(\vect{x}_{n+1})\|} if \|f(\vect{x}_{n+1})\|\neq 0, else \epsilon^r_n=-1.

getMaximumTimeDuration()

Accessor to the maximum duration.

Returns:
maximumTimefloat

Maximum optimization duration in seconds.

getMetamodelNoise()

Metamodel noise model accessor.

Returns:
noiseModelFunction

The noise variance function \sigma^2_{\epsilon}(x) used for the AEI design update only. Of same input dimension as the objective and 1-d output.

getMultiStartExperimentSize()

Size of the design to draw starting points.

Returns:
multiStartExperimentSizeint

The size of the Monte Carlo design from which to select the best starting points.

getMultiStartNumber()

Number of starting points for the criterion optimization.

Returns:
multiStartNumberint

The number of starting points for the criterion optimization.

getName()

Accessor to the object’s name.

Returns:
namestr

The name of the object.

getNoiseModel()

Improvement noise model accessor.

Returns:
noiseModelFunction

The noise variance function \sigma^2_{\epsilon}(x) used for the AEI criterion optimization only. Of same input dimension as the objective and 1-d output.

getOptimizationAlgorithm()

Expected improvement solver accessor.

Returns:
solverOptimizationAlgorithm

The solver used to optimize the expected improvement

getParameterEstimationPeriod()

Parameter estimation period accessor.

Returns:
periodint

The number of iterations between covariance parameters re-learn. Default is 1 (each iteration). Can be set to 0 (never).

getProblem()

Accessor to optimization problem.

Returns:
problemOptimizationProblem

Optimization problem.

getResult()

Accessor to optimization result.

Returns:
resultOptimizationResult

Result class.

getStartingPoint()

Accessor to starting point.

Returns:
startingPointPoint

Starting point.

hasName()

Test if the object is named.

Returns:
hasNamebool

True if the name is not empty.

run()

Launch the optimization.

setAEITradeoff(c)

AEI tradeoff constant accessor.

Parameters:
cfloat

Used to define a quantile of the kriging prediction at the design points. u(x)=m_K(x)+c*s_K(x)

setCheckStatus(checkStatus)

Accessor to check status flag.

Parameters:
checkStatusbool

Whether to check the termination status. If set to False, run() will not throw an exception if the algorithm does not fully converge and will allow one to still find a feasible candidate.

setCorrelationLengthFactor(b)

Correlation length stopping criterion factor accessor.

When a correlation length becomes smaller than the minimal distance between design point for a single component that means the model tends to be noisy, and the EGO formulation is not adapted anymore.

Parameters:
bfloat

Used to define a stopping criterion on the minimum correlation length: \theta_i < \frac{\Delta_i^{min}}{b} with \Delta^{min} the minimum distance between design points.

setImprovementFactor(improvementFactor)

Improvement criterion factor accessor.

Parameters:
alphapositive float, default=0.0 (disables the criterion)

Used to define a stopping criterion on the improvement criterion: I_{max} < \alpha |Y_{min}| with I_{max} the current maximum of the improvement and Y_{min} the current optimum.

setMaximumAbsoluteError(maximumAbsoluteError)

Accessor to maximum allowed absolute error.

Parameters:
maximumAbsoluteErrorfloat

Maximum allowed absolute error, where the absolute error is defined by \epsilon^a_n=\|\vect{x}_{n+1}-\vect{x}_n\|_{\infty} where \vect{x}_{n+1} and \vect{x}_n are two consecutive approximations of the optimum.

setMaximumCallsNumber(maximumCallsNumber)

Accessor to maximum allowed number of calls

Parameters:
maximumEvaluationNumberint

Maximum allowed number of direct objective function calls through the () operator. Does not take into account eventual indirect calls through finite difference gradient calls.

setMaximumConstraintError(maximumConstraintError)

Accessor to maximum allowed constraint error.

Parameters:
maximumConstraintErrorfloat

Maximum allowed constraint error, where the constraint error is defined by \gamma_n=\|g(\vect{x}_n)\|_{\infty} where \vect{x}_n is the current approximation of the optimum and g is the function that gathers all the equality and inequality constraints (violated values only)

setMaximumIterationNumber(maximumIterationNumber)

Accessor to maximum allowed number of iterations.

Parameters:
maximumIterationNumberint

Maximum allowed number of iterations.

setMaximumRelativeError(maximumRelativeError)

Accessor to maximum allowed relative error.

Parameters:
maximumRelativeErrorfloat

Maximum allowed relative error, where the relative error is defined by \epsilon^r_n=\epsilon^a_n/\|\vect{x}_{n+1}\|_{\infty} if \|\vect{x}_{n+1}\|_{\infty}\neq 0, else \epsilon^r_n=-1.

setMaximumResidualError(maximumResidualError)

Accessor to maximum allowed residual error.

Parameters:
maximumResidualErrorfloat

Maximum allowed residual error, where the residual error is defined by \epsilon^r_n=\frac{\|f(\vect{x}_{n+1})-f(\vect{x}_{n})\|}{\|f(\vect{x}_{n+1})\|} if \|f(\vect{x}_{n+1})\|\neq 0, else \epsilon^r_n=-1.

setMaximumTimeDuration(maximumTime)

Accessor to the maximum duration.

Parameters:
maximumTimefloat

Maximum optimization duration in seconds.

setMetamodelNoise(metaModelNoise)

Metamodel noise model accessor.

Parameters:
noiseModelFunction

The noise variance function \sigma^2_{\epsilon}(x) used for the AEI design update only. Of same input dimension as the objective and 1-d output.

setMultiStartExperimentSize(multiStartExperimentSize)

Size of the design to draw starting points.

Parameters:
multiStartExperimentSizeint

The size of the Monte Carlo design from which to select the best starting points. The default number can be tweaked with the EfficientGlobalOptimization-DefaultMultiStartExperimentSize key from ResourceMap.

setMultiStartNumber(multiStartNumberSize)

Number of starting points for the criterion optimization.

Parameters:
multiStartNumberint

The number of starting points for the criterion optimization. The default number can be tweaked with the EfficientGlobalOptimization-DefaultMultiStartNumber key from ResourceMap.

setName(name)

Accessor to the object’s name.

Parameters:
namestr

The name of the object.

setNoiseModel(noiseModel)

Improvement noise model accessor.

Parameters:
noiseModelFunction

The noise variance function \sigma^2_{\epsilon}(x) used for the AEI criterion optimization only. Of same input dimension as the objective and 1-d output.

setOptimizationAlgorithm(solver)

Expected improvement solver accessor.

Parameters:
solverOptimizationAlgorithm

The solver used to optimize the expected improvement

setParameterEstimationPeriod(parameterEstimationPeriod)

Parameter estimation period accessor.

Parameters:
periodint

The number of iterations between covariance parameters re-learn. Default is 1 (each iteration). Can be set to 0 (never). The default number can be tweaked with the EfficientGlobalOptimization-DefaultParameterEstimationPeriod key from ResourceMap.

setProblem(problem)

Accessor to optimization problem.

Parameters:
problemOptimizationProblem

Optimization problem.

setProgressCallback(*args)

Set up a progress callback.

Can be used to programmatically report the progress of an optimization.

Parameters:
callbackcallable

Takes a float as argument as percentage of progress.

Examples

>>> import sys
>>> import openturns as ot
>>> rosenbrock = ot.SymbolicFunction(['x1', 'x2'], ['(1-x1)^2+100*(x2-x1^2)^2'])
>>> problem = ot.OptimizationProblem(rosenbrock)
>>> solver = ot.OptimizationAlgorithm(problem)
>>> solver.setStartingPoint([0, 0])
>>> solver.setMaximumResidualError(1.e-3)
>>> solver.setMaximumCallsNumber(10000)
>>> def report_progress(progress):
...     sys.stderr.write('-- progress=' + str(progress) + '%\n')
>>> solver.setProgressCallback(report_progress)
>>> solver.run()
setResult(result)

Accessor to optimization result.

Parameters:
resultOptimizationResult

Result class.

setStartingPoint(startingPoint)

Accessor to starting point.

Parameters:
startingPointPoint

Starting point.

setStopCallback(*args)

Set up a stop callback.

Can be used to programmatically stop an optimization.

Parameters:
callbackcallable

Returns an int deciding whether to stop or continue.

Examples

>>> import openturns as ot
>>> rosenbrock = ot.SymbolicFunction(['x1', 'x2'], ['(1-x1)^2+100*(x2-x1^2)^2'])
>>> problem = ot.OptimizationProblem(rosenbrock)
>>> solver = ot.OptimizationAlgorithm(problem)
>>> solver.setStartingPoint([0, 0])
>>> solver.setMaximumResidualError(1.e-3)
>>> solver.setMaximumCallsNumber(10000)
>>> def ask_stop():
...     return True
>>> solver.setStopCallback(ask_stop)
>>> solver.run()

Examples using the class

EfficientGlobalOptimization examples

EfficientGlobalOptimization examples