SimulationSensitivityAnalysis is not really working with SystemEvent

[adumasphi]

What happened?

When computing a probability with a system event, the sensitivity analysis is really difficult to apply because the history sample seems not to be propagated inside the result object.
Two constructors on the 3 given are not working at all.

How to reproduce the issue?

import openturns as ot

dim = 2
X = ot.RandomVector(ot.Normal(dim))
f1 = ot.MemoizeFunction(ot.SymbolicFunction(['x1', 'x2'], ['x1']))
f2 = ot.MemoizeFunction(ot.SymbolicFunction(['x1', 'x2'], ['x2']))
Y1 = ot.CompositeRandomVector(f1, X)
Y2 = ot.CompositeRandomVector(f2, X)
e1 = ot.ThresholdEvent(Y1, ot.Less(), 0.0)
e2 = ot.ThresholdEvent(Y2, ot.Greater(), 0.0)
event = ot.UnionEvent([e1, e2])

algo = ot.ProbabilitySimulationAlgorithm(event)
algo.setMaximumOuterSampling(150)
algo.setBlockSize(4)
algo.setMaximumCoefficientOfVariation(0.1)
algo.run()
result = algo.getResult()

# First constructor
# get the history from the result event is not possible
f_event = ot.MemoizeFunction(result.getEvent().getImplementation().asComposedEvent().getFunction())
print(f_event.getCacheInput().getSize())  # --> size = 0
# the event output should be the one of the composite function
# --> it is complicated to get the history, I must recompute it using the event
input_sample = f1.getInputHistory()
output_sample = event.asComposedEvent().getFunction()(input_sample)
ssa = ot.SimulationSensitivityAnalysis(event.asComposedEvent(), input_sample, output_sample)

# second constructor not working : empty samples and maybe not ideal
# to use the getImplementation anyway
ssa = ot.SimulationSensitivityAnalysis(result.getEvent().getImplementation().asComposedEvent())

# third constructor not working
ssa = ot.SimulationSensitivityAnalysis(result)

Version

1.22

Operating System

Linux

Installation media

conda

Additional Context

No response

[jschueller]

A solution is to re-assemble a threshold event from the composed function with a memoize on top:

import openturns as ot

dim = 2
X = ot.RandomVector(ot.Normal(dim))
f1 = ot.SymbolicFunction(['x1', 'x2'], ['x1'])
f2 = ot.SymbolicFunction(['x1', 'x2'], ['x2'])
Y1 = ot.CompositeRandomVector(f1, X)
Y2 = ot.CompositeRandomVector(f2, X)
e1 = ot.ThresholdEvent(Y1, ot.Less(), 0.0)
e2 = ot.ThresholdEvent(Y2, ot.Greater(), 0.0)
event = ot.UnionEvent([e1, e2])

# rebuild event with memoize
event = event.asComposedEvent()
fm = ot.MemoizeFunction(event.getFunction())
crv = ot.CompositeRandomVector(fm, event.getAntecedent())
event = ot.ThresholdEvent(crv, event.getOperator(), event.getThreshold())

algo = ot.ProbabilitySimulationAlgorithm(event)
algo.setMaximumOuterSampling(150)
algo.setBlockSize(4)
algo.setMaximumCoefficientOfVariation(0.1)
algo.run()
result = algo.getResult()

x = fm.getInputHistory()
y = fm.getOutputHistory()
ssa = ot.SimulationSensitivityAnalysis(event, x, y)
print(ssa.computeImportanceFactors())